Merge pull request 'Agent sleep mode: empty queue sheds to one downloader, lease poll backs off to 15 min' (#188) from dev into main

Operator-flagged on the deployed .5 build: the autoscaler grew the pool 1→8
against an EMPTY queue (an empty buffer read as 'GPU starving' regardless of
WHY), and every downloader kept polling lease every 10s all night.

- New idle signal straight from the lease results: an empty lease sets _idle,
  any jobs clear it. The occupancy-low branch now distinguishes three cases:
  queue empty → shed to ONE polling downloader; pinned at the bandwidth cap →
  shed toward 3; cap headroom + work flowing → grow.
- Idle lease polls back off exponentially per downloader to
  IDLE_POLL_MAX_SECONDS (15 min) and reset the moment work appears — so an
  idle night costs one HTTP call per 15 min, and new work is noticed within
  at most ~15 min (operator-accepted trade-off).
- UI hint: 'idle — queue empty, lease poll backed off'; /status gains idle.
  Agent build 2026-07-02.6.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CDgx8bQS5YrGRK76v8HUnM

.forgejo/workflows/ci.yml

@@ -27,7 +27,14 @@ jobs:
     steps:
       - uses: actions/checkout@v4
       - name: Ruff lint
-        run: ruff check backend/ tests/ alembic/
+        # agent/ included so the GPU-agent is linted before its image is built
+        # (build.yml only `docker build`s it — this is where it gets checked).
+        run: ruff check backend/ tests/ alembic/ agent/
+      - name: Agent syntax check
+        # The agent's runtime deps (torch/transformers/ultralytics) aren't in the
+        # CI image, so we can't import it — but compileall parses every module,
+        # catching syntax errors before the image build.
+        run: python -m compileall -q agent/fc_agent
 
   backend-lint-and-test:
     runs-on: python-ci

agent/Dockerfile

@@ -1,18 +1,24 @@
 # FabledCurator GPU agent — runs on the desktop with the GPU.
-# CUDA + cuDNN runtime so onnxruntime-gpu can use the card (it needs cuDNN 9 —
+# CUDA 12.9 + cuDNN 9 runtime so onnxruntime-gpu can use the card (it needs
-# the plain -runtime image lacks it: "libcudnn.so.9: cannot open shared object
+# cuDNN 9 — the plain -runtime image lacks it: "libcudnn.so.9: cannot open
-# file"); ffmpeg for video frames.
+# shared object file"); ffmpeg for video frames. Ubuntu 24.04 → Python 3.12.
-FROM nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04
+# Stays on the CUDA-12 / cuDNN-9 line the default onnxruntime-gpu + torch are
+# built against (CUDA 13 has only nascent ONNX Runtime support).
+FROM nvidia/cuda:12.9.2-cudnn-runtime-ubuntu24.04
 
-ENV DEBIAN_FRONTEND=noninteractive PYTHONUNBUFFERED=1
+# PIP_BREAK_SYSTEM_PACKAGES: Ubuntu 24.04 marks its system Python as externally
+# managed (PEP 668), so a global `pip install` errors without this. It's a
+# single-purpose container — we own the whole environment, so installing into
+# the system site-packages is fine (and simplest — no venv on PATH to manage).
+ENV DEBIAN_FRONTEND=noninteractive PYTHONUNBUFFERED=1 PIP_BREAK_SYSTEM_PACKAGES=1
 RUN apt-get update \
     && apt-get install -y --no-install-recommends python3 python3-pip ffmpeg \
     && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
-# torch from the CUDA-12.4 wheel index (matches the base image); its wheels
+# torch from the CUDA-12.4 wheel index; its wheels bundle their own CUDA + cuDNN
-# bundle their own CUDA + cuDNN and coexist with onnxruntime-gpu. Installed
+# so they run on the 12.9 base and coexist with onnxruntime-gpu. Installed first
-# first + separately so the GPU build of torch is deterministic and layer-cached.
+# + separately so the GPU build of torch is deterministic and layer-cached.
 RUN pip3 install --no-cache-dir torch==2.6.0 --index-url https://download.pytorch.org/whl/cu124
 COPY requirements.txt .
 RUN pip3 install --no-cache-dir -r requirements.txt

agent/docker-compose.yml

@@ -34,9 +34,29 @@ services:
       # Resume the worker automatically on container start (survive a reboot /
       # crash-restart while you're away). Set to 0 to require a manual Start.
       AUTO_START: ${AUTO_START:-1}
+      # Autoscale the worker count (throughput hill-climb that finds the sweet
+      # spot + backs off under VRAM pressure). On by default; toggle live in the
+      # control UI. Set to 0 to start in manual mode.
+      AUTO_SCALE: ${AUTO_SCALE:-1}
+      # Aggregate download cap in MB/s (stills + video streams combined), so the
+      # agent can't saturate the desktop's network and wreck browsing — WiFi
+      # especially. 0 = unlimited; tunable live in the control UI.
+      BANDWIDTH_LIMIT_MB_S: ${BANDWIDTH_LIMIT_MB_S:-8}
       # Crop embedder (SigLIP concept bag): float16 keeps VRAM low on a shared
       # desktop GPU; the model itself is announced by the server.
       SIGLIP_DTYPE: ${SIGLIP_DTYPE:-float16}
+      # Crop PROPOSERS (extra YOLO detectors → more/better concept crops). Each
+      # downloads its weights once (cached on the models volume) and self-disables
+      # if the download/load fails. Blank any one to turn it off.
+      #   PERSON_WEIGHTS: general COCO person detector (Western/realistic figures),
+      #     merged with the anime detector. yolo11n.pt (~6 MB, auto-downloaded).
+      #   ANATOMY_WEIGHTS: booru_yolo anime/furry/NSFW components (~40 MB). NB the
+      #     repo states no license — fine for private use. yolov8n_as01.pt is the
+      #     6 MB nano if you want lighter than yolov11m_aa22.pt.
+      #   PANEL_WEIGHTS: mosesb comic-panel detector (Apache-2.0), "hf_repo::file".
+      PERSON_WEIGHTS: ${PERSON_WEIGHTS:-yolo11n.pt}
+      ANATOMY_WEIGHTS: ${ANATOMY_WEIGHTS:-https://github.com/aperveyev/booru_yolo/raw/main/models/yolov11m_aa22.pt}
+      PANEL_WEIGHTS: ${PANEL_WEIGHTS:-mosesb/best-comic-panel-detection::best.pt}
     volumes:
       # Persist the downloaded ONNX models so restarts are fast.
       - fc-agent-models:/models

agent/fc_agent/app.py

@@ -1,21 +1,45 @@
 """FastAPI control surface for the agent (served on localhost).
 
-Start / stop the worker pool, tune the worker count live (trades desktop
+Start / stop the download→GPU pipeline, tune the downloader count live (the
-responsiveness for throughput), and watch GPU load + progress + the server-side
+workload is download-bound, so downloaders are the dial that trades desktop
-queue. Config is env-seeded; the worker count is adjustable here on the fly.
+bandwidth for throughput), and watch GPU load + buffer occupancy + progress +
+the server-side queue. Config is env-seeded; the downloader count is adjustable
+here on the fly (GPU consumers autoscale between 1 and 2 on their own).
 """
+import logging
+
 from fastapi import FastAPI, Request
 from fastapi.responses import HTMLResponse, JSONResponse
 
+from . import logbuf
 from .config import Config
 from .gpu import read_gpu
 from .worker import Worker
 
+log = logging.getLogger("fc_agent.app")
+
+# Bump on every agent change. The page embeds this and /status reports it; the UI
+# warns to reload when they differ — so a stale browser-cached page can't be
+# mistaken for "the new image didn't deploy". (Belt-and-braces with no-store.)
+VERSION = "2026-07-02.6 · sleep mode: an empty queue sheds to one downloader and backs the lease poll off to 15 min"
+
+logbuf.install()
 cfg = Config.from_env()
 worker = Worker(cfg)
 app = FastAPI(title="FabledCurator GPU agent")
 
 
+@app.middleware("http")
+async def _no_store(request, call_next):
+    # The control page is a static string and the status/gpu/logs polls are
+    # live data — never let the browser cache either, or a freshly-pulled agent
+    # image still shows the OLD UI until a hard refresh (operator-flagged
+    # 2026-06-30).
+    resp = await call_next(request)
+    resp.headers["Cache-Control"] = "no-store"
+    return resp
+
+
 @app.on_event("startup")
 def _maybe_autostart() -> None:
     # With AUTO_START set, a container restart (host reboot, or `restart:
@@ -28,17 +52,19 @@ def _maybe_autostart() -> None:
 
 @app.get("/", response_class=HTMLResponse)
 def index() -> str:
-    return _PAGE
+    return _PAGE.replace("__BUILD__", VERSION)
 
 
 @app.post("/start")
 def start():
+    log.info("UI: Start button pressed")   # the press; worker logs the transition
     worker.start()
     return JSONResponse(worker.status())
 
 
 @app.post("/stop")
 def stop():
+    log.info("UI: Stop button pressed")
     worker.stop()
     return JSONResponse(worker.status())
 
@@ -50,85 +76,329 @@ async def concurrency(request: Request):
     return JSONResponse(worker.status())
 
 
+@app.post("/auto")
+async def auto(request: Request):
+    body = await request.json()
+    worker.set_auto(bool(body.get("value", True)))
+    return JSONResponse(worker.status())
+
+
+@app.post("/bandwidth")
+async def bandwidth(request: Request):
+    body = await request.json()
+    worker.set_bandwidth(float(body.get("value", 0)))
+    return JSONResponse(worker.status())
+
+
+@app.get("/gpu")
+def gpu():
+    # GPU meters poll this on their own fast cadence. It only reads local
+    # nvidia-smi — no curator round-trip — so the util/VRAM bars stay live even
+    # when /status is slow waiting on the (sometimes busy) curator queue call.
+    g = read_gpu() or {}
+    us = worker.util_smooth()
+    if us is not None:
+        g["util_smooth"] = round(us, 1)   # autoscaler's EWMA — the UI bar tracks this
+    return JSONResponse(g)
+
+
+@app.get("/logs")
+def logs():
+    return JSONResponse({"lines": list(logbuf.LINES)})
+
+
 @app.get("/status")
 def status():
+    # Pure in-memory read: worker.status() is lock-free and the queue snapshot is
+    # kept fresh by a background poller — NO inline curator call, so this can't
+    # stall the status view when curator is buried under a big backlog.
+    worker.note_ui()          # a browser is watching → keep the queue snapshot warm
     s = worker.status()
     s["fc_url"] = cfg.fc_url
     s["configured"] = bool(cfg.token)
-    s["gpu"] = read_gpu()
+    s["queue"] = worker.latest_queue()
-    try:
+    s["build"] = VERSION
-        s["queue"] = worker.client.queue_status()
-    except Exception:
-        s["queue"] = None
     return JSONResponse(s)
 
 
 _PAGE = """<!doctype html><html><head><meta charset=utf-8>
-<title>FabledCurator GPU agent</title>
+<meta name=viewport content="width=device-width,initial-scale=1">
+<title>FabledCurator · GPU agent</title>
 <style>
- body{font:14px system-ui;margin:2rem;max-width:680px;background:#14171a;color:#e8e8e8}
+ :root{--bg:#0f1216;--panel:#181c22;--panel2:#1e232b;--bd:#2a313b;--fg:#e9edf2;
- h1{font-size:18px} button{font:14px system-ui;padding:.5rem 1rem;border:0;border-radius:6px;
+  --mut:#8b97a6;--acc:#e8923a;--grn:#46c46a;--red:#e8584d;--amb:#e8b23a}
- margin-right:.5rem;cursor:pointer;color:#fff} .start{background:#2e7d32}.stop{background:#b3261e}
+ *{box-sizing:border-box}
- .step{background:#33373b;padding:.4rem .7rem;font-weight:700}
+ body{font:14px/1.5 system-ui,-apple-system,Segoe UI,Roboto,sans-serif;margin:0;
- .stat{display:inline-block;margin-right:1.5rem;vertical-align:top}
+  background:radial-gradient(1200px 600px at 50% -10%,#1a2029,#0f1216);color:var(--fg)}
- .n{font-size:22px;font-weight:700} code{background:#222;padding:2px 6px;border-radius:4px}
+ .wrap{max-width:820px;margin:0 auto;padding:28px 20px 28px;height:100vh;
- .q,.gpu{margin-top:1rem;color:#9aa} .bar{height:8px;border-radius:4px;background:#222;overflow:hidden;
+  box-sizing:border-box;overflow:hidden;display:flex;flex-direction:column}
- max-width:320px;margin-top:4px} .bar>i{display:block;height:100%;background:#3f7d3f}
+ header{display:flex;align-items:center;justify-content:space-between;margin-bottom:4px}
- .row{margin:.8rem 0}
+ .brand{display:flex;align-items:center;gap:10px;font-size:19px;font-weight:700;letter-spacing:.2px}
+ .logo{color:var(--acc);font-size:20px}
+ .brand .sub{color:var(--mut);font-weight:600;font-size:13px;text-transform:uppercase;letter-spacing:.12em}
+ .conn{display:flex;align-items:center;gap:8px;color:var(--mut);font-size:13px;font-weight:600}
+ .dot{width:9px;height:9px;border-radius:50%;background:var(--mut);box-shadow:0 0 0 0 rgba(0,0,0,0)}
+ .dot.green{background:var(--grn);box-shadow:0 0 10px 1px rgba(70,196,106,.5)}
+ .dot.amber{background:var(--amb)} .dot.red{background:var(--red)}
+ .meta{color:var(--mut);margin:0 0 18px;font-size:13px}
+ code{background:#11151a;border:1px solid var(--bd);padding:2px 7px;border-radius:6px;
+  font:12px ui-monospace,SFMono-Regular,Menlo,monospace;color:#cdd6e0}
+ .card{background:linear-gradient(180deg,var(--panel),var(--panel2));border:1px solid var(--bd);
+  border-radius:14px;padding:16px 18px;margin-bottom:14px;box-shadow:0 1px 0 rgba(255,255,255,.02) inset}
+ .card-h{font-size:11px;font-weight:800;letter-spacing:.12em;text-transform:uppercase;
+  color:var(--mut);margin-bottom:14px}
+ .controls{display:flex;align-items:center;gap:10px;flex-wrap:wrap}
+ .spacer{flex:1}
+ .btn{font:600 14px system-ui;padding:.5rem 1rem;border:1px solid transparent;border-radius:9px;
+  cursor:pointer;color:#fff;transition:.12s}
+ .btn:hover{transform:translateY(-1px)}
+ .btn[disabled]{opacity:.45;pointer-events:none;transform:none}
+ @keyframes pulse{0%,100%{opacity:1}50%{opacity:.4}}
+ .tile .n.busy{color:var(--acc);animation:pulse 1s ease-in-out infinite}
+ .btn.start{background:linear-gradient(180deg,#2f9c4c,#247a3c)}
+ .btn.stop{background:linear-gradient(180deg,#3a3f48,#2a2f37);color:#e9edf2;border-color:var(--bd)}
+ .switch{display:inline-flex;align-items:center;gap:8px;cursor:pointer;font-weight:600;user-select:none}
+ .switch input{display:none}
+ .switch .track{width:38px;height:22px;border-radius:11px;background:#2a313b;position:relative;transition:.15s}
+ .switch .track:after{content:"";position:absolute;top:2px;left:2px;width:18px;height:18px;border-radius:50%;
+  background:#cdd6e0;transition:.15s}
+ .switch input:checked+.track{background:var(--acc)}
+ .switch input:checked+.track:after{transform:translateX(16px);background:#fff}
+ .stepper{display:inline-flex;align-items:center;gap:6px}
+ .step{background:#262c34;color:var(--fg);border:1px solid var(--bd);border-radius:8px;
+  width:30px;height:32px;font:700 16px system-ui;cursor:pointer}
+ .step:hover{border-color:var(--acc)}
+ #conc,#bw{width:3.4rem;height:32px;text-align:center;font:700 16px system-ui;background:#11151a;
+  color:var(--fg);border:1px solid var(--bd);border-radius:8px}
+ .unit{color:var(--mut);font-size:12px;font-weight:600}
+ .hint{color:var(--mut);font-size:12px;margin-top:12px}
+ .tiles{display:grid;grid-template-columns:repeat(6,1fr);gap:8px;margin-bottom:16px}
+ .tile{background:#13171d;border:1px solid var(--bd);border-radius:10px;padding:12px 8px;text-align:center}
+ .tile .n{font:800 22px ui-monospace,monospace;line-height:1.1}
+ .tile .n.warn{color:var(--red)} .tile .n.ok{color:var(--grn)}
+ .tile .l{font-size:10px;text-transform:uppercase;letter-spacing:.06em;color:var(--mut);margin-top:4px}
+ .meters{display:flex;flex-direction:column;gap:10px;margin-bottom:14px}
+ .meter-h{display:flex;justify-content:space-between;font-size:12px;color:var(--mut);margin-bottom:4px}
+ .meter-h b{color:var(--fg);font-variant-numeric:tabular-nums}
+ .bar{height:9px;border-radius:5px;background:#11151a;border:1px solid var(--bd);overflow:hidden}
+ .bar>i{display:block;height:100%;width:0;background:linear-gradient(90deg,#3a7d57,var(--grn));transition:width .4s}
+ #utilbar{background:linear-gradient(90deg,#9a5a1f,var(--acc))}
+ #bufbar{background:linear-gradient(90deg,#2f5a9a,#4a86d8)}
+ .queue{font:13px ui-monospace,monospace;color:var(--mut)}
+ .banner{margin:0 0 14px;padding:.7rem .9rem;border-radius:10px;background:#3a2f12;
+  border:1px solid #5a4a17;color:#ffd98a;font-size:13px}
+ .logs-h{display:flex;align-items:center;justify-content:space-between}
+ .grow{flex:1;display:flex;flex-direction:column;min-height:0}
+ .grow .logs{flex:1;min-height:0}
+ .copybtn{font:600 11px system-ui;letter-spacing:.04em;text-transform:uppercase;
+  background:#262c34;color:var(--fg);border:1px solid var(--bd);border-radius:7px;
+  padding:5px 11px;cursor:pointer}
+ .copybtn:hover{border-color:var(--acc)}
+ .logs{margin:0;background:#0b0e12;border:1px solid var(--bd);border-radius:10px;padding:12px;
+  overflow:auto;font:12px/1.55 ui-monospace,SFMono-Regular,Menlo,monospace;
+  color:#b9c4d0;white-space:pre-wrap;word-break:break-word}
 </style></head><body>
-<h1>FabledCurator GPU agent</h1>
+<div class=wrap>
-<p>FC: <code id=fc>—</code> · token <code id=cfg>—</code></p>
+ <header>
-<div class=row>
+  <div class=brand><span class=logo>◆</span> FabledCurator <span class=sub>GPU agent</span></div>
- <button class=start onclick=act('start')>Start</button>
+  <div class=conn><span class="dot" id=dot></span><span id=connlbl>—</span></div>
- <button class=stop onclick=act('stop')>Stop</button>
+ </header>
+ <p class=meta>Server <code id=fc>—</code> · token <code id=cfg>—</code> · build <code id=build>__BUILD__</code></p>
+
+ <div id=verbanner class=banner style="display:none;background:#3a1212;border-color:#5a1717;color:#ffb3b3">
+  a newer agent version is running — reload this page (Ctrl+Shift+R) to update the controls
+ </div>
+ <div id=banner class=banner style=display:none>
+  curator unreachable — holding work + retrying, resumes on its own (no restart needed)
+ </div>
+
+ <section class=card>
+  <div class=card-h>Control</div>
+  <div class=controls>
+   <button class="btn start" id=startbtn onclick=act('start')>▶ Start</button>
+   <button class="btn stop" id=stopbtn onclick=act('stop')>■ Stop</button>
+   <div class=spacer></div>
+   <label class=switch><input type=checkbox id=autochk onchange="setauto(this.checked)"><span class=track></span>Auto</label>
+   <div class=stepper>
+    <button class=step onclick=setc(-1)>−</button>
+    <input id=conc type=number min=1 value=1 onchange="setv(this.value)">
+    <button class=step onclick=setc(1)>+</button>
+   </div>
+   <div class=stepper title="aggregate download cap, downloads + video streams combined — 0 = unlimited">
+    <input id=bw type=number min=0 step=1 value=8 onchange="setbw(this.value)">
+    <span class=unit>MB/s</span>
+   </div>
+  </div>
+  <div class=hint id=conchint>auto-tuning downloaders to keep the GPU fed · max 8</div>
+ </section>
+
+ <section class=card>
+  <div class=card-h>Status</div>
+  <div class=tiles>
+   <div class=tile><div class=n id=state>—</div><div class=l>state</div></div>
+   <div class=tile><div class=n id=jpm>—</div><div class=l>jobs / min</div></div>
+   <div class=tile><div class=n id=dpm>—</div><div class=l>downloads / min</div></div>
+   <div class=tile><div class="n ok" id=done>0</div><div class=l>processed</div></div>
+   <div class=tile><div class=n id=err>0</div><div class=l>errors</div></div>
+   <div class=tile><div class=n id=waited>0</div><div class=l>waited out</div></div>
+  </div>
+  <div class=meters>
+   <div class=meter><div class=meter-h><span>GPU util</span><b id=utillbl>—</b></div>
+    <div class=bar><i id=utilbar></i></div></div>
+   <div class=meter><div class=meter-h><span>VRAM</span><b id=vramlbl>—</b></div>
+    <div class=bar><i id=gpubar></i></div></div>
+   <div class=meter><div class=meter-h><span>buffer occupancy</span><b id=buflbl>—</b></div>
+    <div class=bar><i id=bufbar></i></div></div>
+  </div>
+  <div class=queue id=pipe>downloaders — · consumers — · on GPU 0</div>
+  <div class=queue id=queue>queue —</div>
+ </section>
+
+ <section class="card grow">
+  <div class="card-h logs-h">Logs
+   <button class=copybtn id=copybtn onclick=copyLogs()>Copy</button>
+  </div>
+  <pre class=logs id=logs>waiting for activity…</pre>
+ </section>
 </div>
-<div class=row>
- workers
- <button class=step onclick=setc(-1)>−</button>
- <input id=conc type=number min=1 value=1
-  style="width:3.5rem;font:700 16px system-ui;text-align:center;background:#222;color:#e8e8e8;border:1px solid #444;border-radius:6px;padding:.3rem"
-  onchange="setv(this.value)">
- <button class=step onclick=setc(1)>+</button>
- <span class=cap style=color:#9aa>(more = overlap I/O, fill the GPU) max <b id=capn>8</b></span>
-</div>
-<div class=row>
- <span class=stat><span class=n id=state>stopped</span><br>state</span>
- <span class=stat><span class=n id=active>0</span><br>active now</span>
- <span class=stat><span class=n id=done>0</span><br>processed</span>
- <span class=stat><span class=n id=err>0</span><br>errors</span>
- <span class=stat><span class=n id=wait>0</span><br>waited out</span>
-</div>
-<div id=banner style="display:none;margin:.6rem 0;padding:.5rem .8rem;border-radius:6px;background:#5a4a17;color:#ffe28a">
- curator unreachable — holding work + retrying, will resume on its own (no restart needed)
-</div>
-<div class=gpu id=gpu>GPU — …</div>
-<div class=bar><i id=gpubar style=width:0%></i></div>
-<div class=q id=queue></div>
 <script>
+ const PAGE_BUILD="__BUILD__"
  let CAP=8
- async function act(p){await fetch('/'+p,{method:'POST'});refresh()}
+ // Optimistic transitional state on click, then apply the POST's own status
- function setc(d){ setv((parseInt(conc.value||'1'))+d) }
+ // response (it returns worker.status()) for instant feedback — don't wait on the
+ // separate /status poll, which can lag behind the curator queue call.
+ async function act(p){
+  pending(p==='start'?'starting':'stopping')
+  // Abort a slow POST after 8s so the buttons never stay stuck — the periodic
+  // /status refresh (now always fast) recovers the true state either way.
+  const ac=new AbortController(); const to=setTimeout(()=>ac.abort(),8000)
+  try{ applyStatus(await (await fetch('/'+p,{method:'POST',signal:ac.signal})).json()) }
+  catch{ refresh() /* on abort/error, repaint the real state from /status */ }
+  finally{ clearTimeout(to) }
+ }
+ function pending(label){
+  // Instant optimistic feedback on click; applyStatus (POST response, then the
+  // periodic poll) then owns the real state + which buttons are enabled.
+  state.textContent=label; state.className='n busy'
+  dot.className='dot amber'
+  startbtn.disabled=true; stopbtn.disabled=true
+ }
+ function setc(d){ if(conc.disabled)return; setv((parseInt(conc.value||'1'))+d) }
  async function setv(v){
   v=Math.max(1,Math.min(CAP,parseInt(v)||1)); conc.value=v
   await fetch('/concurrency',{method:'POST',headers:{'Content-Type':'application/json'},
    body:JSON.stringify({value:v})});refresh()
  }
+ async function setauto(on){
+  await fetch('/auto',{method:'POST',headers:{'Content-Type':'application/json'},
+   body:JSON.stringify({value:on})});refresh()
+ }
+ async function setbw(v){
+  v=Math.max(0,parseFloat(v)||0); bw.value=v
+  await fetch('/bandwidth',{method:'POST',headers:{'Content-Type':'application/json'},
+   body:JSON.stringify({value:v})});refresh()
+ }
  async function refresh(){
-  const s=await (await fetch('/status')).json()
+  let s; try{ s=await (await fetch('/status')).json() }catch{ return }
-  CAP=s.max_concurrency||8; capn.textContent=CAP
+  applyStatus(s)
-  state.textContent=s.state; active.textContent=s.active; done.textContent=s.processed
+ }
-  err.textContent=s.errors; fc.textContent=s.fc_url; wait.textContent=s.transient||0
+ function applyStatus(s){
-  // Running but the queue read failed → curator is unreachable; show we're
+  // NB: don't write a separate `capn` element here — conchint.textContent below
-  // riding it out rather than erroring.
+  // rewrites the whole hint (incl. the max), and any child element nested in it
-  banner.style.display=(s.state==='running' && !s.queue)?'block':'none'
+  // would be destroyed by that write, breaking the NEXT applyStatus call.
+  CAP=s.max_concurrency||8
+  // The backend owns the state now (stopped|starting|running|stopping) and drives
+  // every transition, so the pill is always truthful — no client-side guessing
+  // from active>0, which used to wedge on "stopping" forever.
+  const st=s.state||'stopped'
+  const running=st==='running'
+  const busy=(st==='starting'||st==='stopping')
+  // Stale-page guard: if the server is a newer build than this page, the cached
+  // controls may misbehave — tell the operator to reload.
+  if(s.build && s.build!==PAGE_BUILD) verbanner.style.display='block'
+  state.textContent=st
+  state.className='n'+(busy?' busy':'')
+  // Buttons follow the real state so you can't fight a transition: Start only
+  // from stopped; Stop only while up; both disabled through "stopping" until the
+  // backend truthfully lands on "stopped".
+  startbtn.disabled=(st!=='stopped')
+  stopbtn.disabled=!(running||st==='starting')
+  // Throughput rates arrive READY from the backend (jobs/min ≈ GPU throughput,
+  // dl/min ≈ fetch throughput), computed there on a fixed cadence — so they show
+  // a real number no matter how often this tab polls (a backgrounded tab throttles
+  // its timers, which used to leave a client-side delta-rate blank forever).
+  jpm.textContent=(s.jobs_per_min!=null)?Math.round(s.jobs_per_min):'—'
+  dpm.textContent=(s.downloads_per_min!=null)?Math.round(s.downloads_per_min):'—'
+  done.textContent=s.processed
+  err.textContent=s.errors; err.className='n'+(s.errors>0?' warn':'')
+  waited.textContent=s.transient||0
+  // Instantaneous pool state → demoted to the sub-line, where its jumpiness reads
+  // as live churn rather than a "broken" headline metric.
+  pipe.textContent='downloaders '+(s.downloaders!=null?s.downloaders:'—')+' · consumers '+(s.consumers!=null?s.consumers:'—')+' · on GPU '+(s.active||0)
+   +' · net '+(s.net_mb_s!=null?s.net_mb_s.toFixed(1):'—')+' MB/s'
+   +(s.bandwidth_limit_mb_s>0?(' / cap '+s.bandwidth_limit_mb_s):'')
+  if(document.activeElement!==bw && s.bandwidth_limit_mb_s!=null) bw.value=s.bandwidth_limit_mb_s
+  // Buffer occupancy bar (also driven here so it tracks the /status cadence).
+  if(s.buffer!=null && s.buffer_max){ const p=Math.round(100*s.buffer/s.buffer_max)
+   buflbl.textContent=s.buffer+' / '+s.buffer_max; bufbar.style.width=p+'%' }
+  // Auto on → dial reflects the auto-chosen count (read-only); off → manual.
+  if(document.activeElement!==autochk) autochk.checked=!!s.auto
+  conc.disabled=!!s.auto; conc.style.opacity=s.auto?0.55:1
+  conchint.textContent=(s.auto?('auto-tuning downloaders to keep the GPU fed · max '+CAP):('manual downloaders · max '+CAP))
+   +(s.idle?' · idle — queue empty, lease poll backed off (new work noticed within ~15 min)'
+     :(s.bw_capped?' · holding at the bandwidth cap (more downloaders would not go faster)':''))
   if(document.activeElement!==conc) conc.value=s.concurrency
   conc.max=CAP
-  cfg.textContent=s.configured?'set':'MISSING'
+  // Connection pill + queue come only from the /status poll (the Start/Stop POST
-  if(s.gpu){
+  // responses skip the slow curator call to stay snappy) — guard so an action
-   gpu.textContent=`GPU — ${s.gpu.util_pct}% util · VRAM ${s.gpu.mem_used_mb}/${s.gpu.mem_total_mb} MB · ${s.gpu.temp_c}°C`
+  // response doesn't blank them.
-   gpubar.style.width=Math.round(100*s.gpu.mem_used_mb/s.gpu.mem_total_mb)+'%'
+  if('configured' in s){
-  } else { gpu.textContent='GPU — n/a (CPU fallback?)'; gpubar.style.width='0%' }
+   const ok=s.configured
-  queue.textContent=s.queue?`queue — pending ${s.queue.pending} · in flight ${s.queue.leased} · done ${s.queue.done} · errored ${s.queue.error}`:'queue — unreachable'
+   fc.textContent=s.fc_url; cfg.textContent=ok?'set':'MISSING'
+   // Pill colour + label track the real state: green only when running AND
+   // curator is answering; amber for the transient states + a running-but-
+   // unreachable curator; grey when stopped; red with no token.
+   let dc='dot', lbl='stopped'
+   if(!ok){ dc='dot red'; lbl='no token' }
+   else if(st==='running'){ dc='dot '+(s.queue?'green':'amber'); lbl=s.queue?'running':'running · curator unreachable' }
+   else if(st==='starting'){ dc='dot amber'; lbl='starting…' }
+   else if(st==='stopping'){ dc='dot amber'; lbl='stopping…' }
+   dot.className=dc; connlbl.textContent=lbl
+   banner.style.display=(st==='running' && !s.queue)?'block':'none'
+   queue.textContent=s.queue?('queue · pending '+s.queue.pending+' · in flight '+s.queue.leased+' · done '+s.queue.done+' · errored '+s.queue.error):'queue · unreachable'
+  }
  }
- refresh(); setInterval(refresh,3000)
+ // GPU meters poll their OWN endpoint on a fast cadence — kept off /status so a
+ // slow curator queue call can't freeze the bars (they only stale on refresh).
+ let UAVG=null   // smoothed util for the bar (raw util swings 0↔99; show the trend)
+ async function refreshGpu(){
+  let g; try{ g=await (await fetch('/gpu')).json() }catch{ return }
+  if(g && g.util_pct!=null){
+   // Prefer the agent's own EWMA (util_smooth) when running; otherwise smooth
+   // the raw reading here so a stopped agent's bar still glides, not jumps.
+   const raw=g.util_pct
+   UAVG = (g.util_smooth!=null) ? g.util_smooth
+         : (UAVG==null ? raw : 0.25*raw + 0.75*UAVG)
+   const used=g.mem_used_mb, tot=g.mem_total_mb||1
+   utillbl.textContent=Math.round(UAVG)+'% · '+g.temp_c+'°C'; utilbar.style.width=Math.round(UAVG)+'%'
+   vramlbl.textContent=used+' / '+tot+' MB'; gpubar.style.width=Math.round(100*used/tot)+'%'
+  } else { UAVG=null; utillbl.textContent='n/a'; vramlbl.textContent='n/a (CPU?)'; utilbar.style.width='0%'; gpubar.style.width='0%' }
+ }
+ async function refreshLogs(){
+  try{
+   const r=await (await fetch('/logs')).json()
+   const el=logs, atBottom=el.scrollHeight-el.scrollTop-el.clientHeight<40
+   el.textContent=(r.lines&&r.lines.length)?r.lines.join('\\n'):'waiting for activity…'
+   if(atBottom) el.scrollTop=el.scrollHeight
+  }catch{}
+ }
+ async function copyLogs(){
+  const txt=logs.textContent||''
+  try{ await navigator.clipboard.writeText(txt) }
+  catch{ const t=document.createElement('textarea'); t.value=txt; document.body.appendChild(t);
+   t.select(); try{document.execCommand('copy')}catch{}; t.remove() }
+  copybtn.textContent='Copied'; setTimeout(()=>{copybtn.textContent='Copy'},1200)
+ }
+ refresh(); refreshGpu(); refreshLogs()
+ setInterval(refresh,3000); setInterval(refreshGpu,1500); setInterval(refreshLogs,2500)
 </script></body></html>"""

agent/fc_agent/client.py

@@ -5,19 +5,69 @@ bytes, all over HTTP with the bearer token. No DB/Redis.
 """
 import requests
 from requests.adapters import HTTPAdapter
+from urllib3.util.retry import Retry
 
 
 class FcClient:
     def __init__(self, base_url: str, token: str, agent_id: str):
         self.base = base_url.rstrip("/")
         self.agent_id = agent_id
-        self.s = requests.Session()
+        # Main session: NO in-request retry — lease/fetch are cheap to redo and
-        self.s.headers["Authorization"] = f"Bearer {token}"
+        # the worker loop already backs off + re-leases on failure. (Auto-retrying
-        # Many worker threads share this Session; the default pool (10) would
+        # a lease could double-claim a batch if a response is lost.)
+        self.s = self._session(token)
+        # Submit session: retry in-place, because by submit time the GPU work is
+        # already DONE — a momentary blip (dropped connection, gateway 5xx during
+        # a curator redeploy) must not throw that work away and force a full
+        # re-download + recompute on another agent. A duplicate submit after a
+        # lost response is harmless: the job is already closed, so it just returns
+        # 409 lease_invalid (a no-op). Idempotent enough to retry POST safely.
+        retry = Retry(
+            total=3, connect=3, read=3, status=3,
+            backoff_factor=0.5,                       # ~0.5s, 1s, 2s between tries
+            status_forcelist=(500, 502, 503, 504),    # transient server/gateway
+            allowed_methods=frozenset({"POST"}),
+            raise_on_status=False,                     # let raise_for_status decide
+        )
+        self._submit_s = self._session(token, retry)
+
+    @staticmethod
+    def _session(token: str, retry: Retry | None = None) -> requests.Session:
+        s = requests.Session()
+        s.headers["Authorization"] = f"Bearer {token}"
+        # Many worker threads share a Session; the default pool (10) would
         # throttle them + spam "connection pool is full". Size it for the cap.
-        adapter = HTTPAdapter(pool_connections=64, pool_maxsize=64)
+        adapter = HTTPAdapter(
-        self.s.mount("http://", adapter)
+            pool_connections=64, pool_maxsize=64, max_retries=retry or 0
-        self.s.mount("https://", adapter)
+        )
+        s.mount("http://", adapter)
+        s.mount("https://", adapter)
+        return s
+
+    def _submit(self, path: str, payload: dict) -> dict:
+        """POST to a submit endpoint on the RETRYING session (by submit time the
+        GPU work is done — a blip must not throw it away), raise on a hard error,
+        and return the parsed JSON. `agent_id` is added to every body."""
+        r = self._submit_s.post(
+            f"{self.base}{path}",
+            json={"agent_id": self.agent_id, **payload},
+            timeout=120,
+        )
+        r.raise_for_status()
+        return r.json()
+
+    def _post_quiet(self, path: str, payload: dict) -> None:
+        """Fire-and-forget POST on the main session — heartbeat/fail/release are
+        best-effort, so a transport error is swallowed (the worker's own retry and
+        the server's orphan-recovery cover a lost call). `agent_id` is added."""
+        try:
+            self.s.post(
+                f"{self.base}{path}",
+                json={"agent_id": self.agent_id, **payload},
+                timeout=30,
+            )
+        except requests.RequestException:
+            pass
 
     def lease(self, batch_size: int) -> list[dict]:
         r = self.s.post(
@@ -29,57 +79,62 @@ class FcClient:
         return r.json().get("jobs", [])
 
     def submit(self, job_id: int, regions: list[dict], replace_kinds: list[str]) -> dict:
-        r = self.s.post(
+        return self._submit("/api/gpu/jobs/submit", {
-            f"{self.base}/api/gpu/jobs/submit",
+            "job_id": job_id, "regions": regions, "replace_kinds": replace_kinds,
-            json={
+        })
-                "agent_id": self.agent_id, "job_id": job_id,
+
-                "regions": regions, "replace_kinds": replace_kinds,
+    def submit_embedding(self, job_id: int, embedding: list, version: str) -> dict:
-            },
+        """Post a whole-image SigLIP embedding (the 'embed' task) → image_record."""
-            timeout=120,
+        return self._submit("/api/gpu/jobs/submit_embedding", {
-        )
+            "job_id": job_id, "embedding": embedding, "embedding_version": version,
-        r.raise_for_status()
+        })
-        return r.json()
 
     def heartbeat(self, job_ids: list[int]) -> None:
-        try:
+        self._post_quiet("/api/gpu/jobs/heartbeat", {"job_ids": job_ids})
-            self.s.post(
-                f"{self.base}/api/gpu/jobs/heartbeat",
-                json={"agent_id": self.agent_id, "job_ids": job_ids},
-                timeout=30,
-            )
-        except requests.RequestException:
-            pass
 
     def fail(self, job_id: int, error: str) -> None:
-        try:
+        self._post_quiet("/api/gpu/jobs/fail", {"job_id": job_id, "error": error})
-            self.s.post(
-                f"{self.base}/api/gpu/jobs/fail",
-                json={"agent_id": self.agent_id, "job_id": job_id, "error": error},
-                timeout=30,
-            )
-        except requests.RequestException:
-            pass
 
     def release(self, job_ids: list[int]) -> None:
         # Graceful hand-back on stop so orphaned work is re-leased at once.
         if not job_ids:
             return
-        try:
+        self._post_quiet("/api/gpu/jobs/release", {"job_ids": job_ids})
-            self.s.post(
-                f"{self.base}/api/gpu/jobs/release",
-                json={"agent_id": self.agent_id, "job_ids": job_ids},
-                timeout=30,
-            )
-        except requests.RequestException:
-            pass
 
-    def fetch_image(self, image_url: str) -> bytes:
+    def fetch_image(self, image_url: str, throttle=None) -> bytes:
         # image_url is a server-relative path ("/images/...").
-        r = self.s.get(f"{self.base}{image_url}", timeout=180)
+        # timeout=(connect, read): the read timeout is BETWEEN-BYTES, not total,
-        r.raise_for_status()
+        # so a large-but-flowing download still completes — but a stuck/dead
-        return r.content
+        # connection (curator overloaded) fails in 60s instead of hanging a
+        # downloader for 180s and piling up concurrent stuck requests on curator.
+        # With a throttle (the worker's shared TokenBucket), the body is streamed
+        # in chunks and each chunk is charged to the global bandwidth budget —
+        # pausing between reads lets TCP flow control pace curator's send side.
+        with self.s.get(
+            f"{self.base}{image_url}", timeout=(10, 60), stream=throttle is not None
+        ) as r:
+            r.raise_for_status()
+            if throttle is None:
+                return r.content
+            buf = bytearray()
+            for chunk in r.iter_content(chunk_size=262_144):
+                throttle.take(len(chunk))
+                buf.extend(chunk)
+            return bytes(buf)
+
+    def is_reachable(self) -> bool:
+        """Cheap 'is curator responding at all right now?' check. Used to decide,
+        when a video can't be sampled, between a transient outage (keep retrying —
+        survives a redeploy) and an unprocessable file (fail it, don't loop)."""
+        try:
+            r = self.s.get(f"{self.base}/api/gpu/status", timeout=5)
+            return r.status_code < 500
+        except requests.RequestException:
+            return False
 
     def queue_status(self) -> dict:
-        r = self.s.get(f"{self.base}/api/gpu/status", timeout=15)
+        # Short timeout: this backs the UI /status poll, so a busy curator must
+        # not hang the page for long (the GPU meters poll /gpu separately).
+        r = self.s.get(f"{self.base}/api/gpu/status", timeout=5)
         r.raise_for_status()
         return r.json()

agent/fc_agent/config.py

@@ -1,8 +1,18 @@
 """Agent config, all from env (the control container is configured at run)."""
+# Lazy annotations so the `from_env(cls) -> Config` self-reference is a string,
+# not evaluated at class-definition time — otherwise it NameErrors on the agent's
+# Python 3.10 (CI lints on 3.14, where PEP 649 hides this).
+from __future__ import annotations
+
 import os
 from dataclasses import dataclass
 
 
+def _bool_env(name: str, default: str = "") -> bool:
+    """A boolean env var — present + truthy ('1'/'true'/'yes') → True."""
+    return os.environ.get(name, default).lower() in ("1", "true", "yes")
+
+
 @dataclass
 class Config:
     fc_url: str          # base URL of the FabledCurator web service
@@ -18,9 +28,32 @@ class Config:
     #                            the server announces in the lease)
     auto_start: bool     # start the worker pool on boot (so a container restart
     #                      resumes processing without anyone clicking Start)
+    auto_scale: bool     # autoscale the worker count (throughput hill-climb)
+    # Crop PROPOSERS (extra YOLO detectors that say where to crop). Each weight
+    # spec is an ultralytics name | http(s) URL | "hf_repo::file" ("" = off).
+    person_weights: str  # general COCO person detector (Western/realistic figs)
+    person_conf: float
+    anatomy_weights: str  # booru_yolo anime/furry/NSFW components
+    anatomy_conf: float
+    panel_weights: str    # comic-panel detector
+    panel_conf: float
+    max_components: int   # cap anatomy component crops per frame
+    max_panels: int       # cap panel crops per frame
+    max_figures: int      # cap figure boxes per frame (each = a CCIP call + crop)
+    max_regions: int      # hard cap on total regions per JOB (submit-size backstop)
+    dedupe_iou: float     # crops overlapping >= this (same kind) are near-dupes,
+    #                       dropped before the embed; >=1.0 disables it
+    frame_dedupe_distance: int  # video frames whose dHash differs by < this many
+    #                             bits are near-duplicates, dropped before detect;
+    #                             higher keeps more frames, 0 disables
+    ffmpeg_timeout: float  # hard ceiling (s) for ffmpeg-from-URL video sampling;
+    #                        generous so a SLOW media link still completes
+    bandwidth_limit_mb_s: float  # aggregate download cap in MEGABYTES/s across
+    #                              all downloaders + video streams (0 = unlimited);
+    #                              tunable live from the agent UI
 
     @classmethod
-    def from_env(cls) -> "Config":
+    def from_env(cls) -> Config:
         return cls(
             fc_url=os.environ.get("FC_URL", "http://localhost:8000").rstrip("/"),
             token=os.environ.get("FC_TOKEN", ""),
@@ -32,5 +65,26 @@ class Config:
             poll_idle_seconds=float(os.environ.get("POLL_IDLE_SECONDS", "10")),
             embed_dtype=os.environ.get("SIGLIP_DTYPE", "float16"),
             embed_model_override=os.environ.get("EMBED_MODEL_NAME", ""),
-            auto_start=os.environ.get("AUTO_START", "").lower() in ("1", "true", "yes"),
+            auto_start=_bool_env("AUTO_START"),
+            auto_scale=_bool_env("AUTO_SCALE", "true"),
+            person_weights=os.environ.get("PERSON_WEIGHTS", "yolo11n.pt"),
+            person_conf=float(os.environ.get("PERSON_CONF", "0.35")),
+            anatomy_weights=os.environ.get("ANATOMY_WEIGHTS", ""),
+            anatomy_conf=float(os.environ.get("ANATOMY_CONF", "0.30")),
+            panel_weights=os.environ.get("PANEL_WEIGHTS", ""),
+            panel_conf=float(os.environ.get("PANEL_CONF", "0.30")),
+            max_components=int(os.environ.get("MAX_COMPONENTS", "8")),
+            max_panels=int(os.environ.get("MAX_PANELS", "8")),
+            max_figures=int(os.environ.get("MAX_FIGURES", "8")),
+            max_regions=int(os.environ.get("MAX_REGIONS", "128")),
+            dedupe_iou=float(os.environ.get("DEDUPE_IOU", "0.85")),
+            frame_dedupe_distance=int(os.environ.get("FRAME_DEDUPE_DISTANCE", "8")),
+            ffmpeg_timeout=float(os.environ.get("FFMPEG_TIMEOUT", "1200")),
+            # Default 8 MB/s (~64 Mbit/s): ~20% of the measured ~300 Mbit/s home
+            # WiFi, so browsing stays snappy while the agent works — yet MORE
+            # sweep throughput than the self-inflicted congestion collapse this
+            # replaces (2026-07-02: 8 unthrottled downloaders bufferbloated the
+            # link to ~1-1.5 MB/s per stream, browser included). Raise it (or 0)
+            # from the agent UI on wired/faster networks.
+            bandwidth_limit_mb_s=float(os.environ.get("BANDWIDTH_LIMIT_MB_S", "8")),
         )

agent/fc_agent/detectors.py

@@ -0,0 +1,218 @@
+"""Region PROPOSERS — small YOLO detectors that decide WHERE to crop. They run
+on the agent GPU and their boxes feed the crop → SigLIP → max-over-bag pipeline:
+
+  - person (general COCO yolo11n): full-figure boxes for realistic / Western art
+    the anime person-detector misses; NMS-merged with imgutils detect_person and
+    fed to CCIP (identity) + a concept crop.
+  - anatomy (booru_yolo): anime / furry / NSFW torso components (head, cat-head,
+    boob, hip, …) — concept crops aligned to the operator's tag vocabulary.
+  - panel (mosesb): a comic page → panel regions → concept crops.
+
+Each proposer is INDEPENDENTLY optional + guarded: a bad weight path or an
+inference error disables just that proposer (logged) and never breaks the
+worker, which still falls back to imgutils detection. Weights resolve from an
+ultralytics builtin name ("yolo11n.pt"), an http(s) URL, or "hf_repo::file" —
+cached under HF_HOME so the download happens once.
+"""
+import logging
+import os
+import threading
+import types
+from pathlib import Path
+
+log = logging.getLogger("fc_agent.detectors")
+_CACHE = Path(os.environ.get("HF_HOME", "/models")) / "yolo"
+
+
+def _resolve(spec: str) -> str | None:
+    """A local weights path (downloading if needed) or an ultralytics builtin
+    name. None if the spec is empty/unresolvable."""
+    if not spec:
+        return None
+    if "::" in spec:                                    # hf_repo::filename
+        repo, _, fname = spec.partition("::")
+        from huggingface_hub import hf_hub_download
+        return hf_hub_download(
+            repo_id=repo, filename=fname, cache_dir=str(_CACHE)
+        )
+    if spec.startswith(("http://", "https://")):
+        _CACHE.mkdir(parents=True, exist_ok=True)
+        dest = _CACHE / spec.rsplit("/", 1)[-1]
+        if not dest.is_file():
+            import requests
+            r = requests.get(spec, timeout=300)
+            r.raise_for_status()
+            dest.write_bytes(r.content)
+        return str(dest)
+    return spec                                         # ultralytics builtin name
+
+
+def _iou(a, b) -> float:
+    ax, ay, aw, ah = a
+    bx, by, bw, bh = b
+    ix = max(0.0, min(ax + aw, bx + bw) - max(ax, bx))
+    iy = max(0.0, min(ay + ah, by + bh) - max(ay, by))
+    inter = ix * iy
+    union = aw * ah + bw * bh - inter
+    return inter / union if union > 0 else 0.0
+
+
+def nms_merge(boxes, iou_thresh: float = 0.6):
+    """Greedy NMS over (bbox_norm, score, label) from possibly several detectors,
+    so the same figure found by two of them collapses to one (higher-score) box."""
+    kept = []
+    for bb, sc, lb in sorted(boxes, key=lambda b: b[1], reverse=True):
+        if all(_iou(bb, k[0]) < iou_thresh for k in kept):
+            kept.append((bb, sc, lb))
+    return kept
+
+
+def dedupe_crops(pending, iou_thresh: float = 0.85):
+    """Greedy high-IoU dedupe over a list of (crop, region_template) pairs, run
+    just before the batched SigLIP embed so we never embed the same region twice.
+
+    Figure boxes are already NMS-merged and each YOLO self-NMSes, but the combined
+    per-frame pile (figure→concept ∪ anatomy component→concept ∪ panel) can still
+    carry genuine near-duplicates across proposers — e.g. a figure box that nearly
+    coincides with an anatomy component on a solo bust, or overlapping booru head
+    classes on one head. Those embed the same region twice, wasting GPU and a slot
+    against max_regions.
+
+    Boxes are compared ONLY within the same output kind and dropped when they
+    overlap at >= iou_thresh, keeping the highest-scoring one. The HIGH default
+    threshold is deliberate: it collapses only true near-identical boxes while
+    preserving intentional nested crops across scopes (a whole figure vs a small
+    head component sit well below it) and distinct kinds (concept vs panel). A
+    value >= 1.0 effectively disables it (nothing but an exact box matches)."""
+    kept = []
+    kept_boxes: dict = {}   # kind -> [bbox, ...] already kept
+    for crop, tmpl in sorted(
+        pending, key=lambda p: p[1].get("score") or 0.0, reverse=True
+    ):
+        bb = tmpl.get("bbox")
+        prior = kept_boxes.setdefault(tmpl.get("kind"), [])
+        if bb is not None and any(_iou(bb, kb) >= iou_thresh for kb in prior):
+            continue
+        prior.append(bb)
+        kept.append((crop, tmpl))
+    return kept
+
+
+class YoloProposer:
+    """One lazily-loaded ultralytics YOLO. detect(image) → [(bbox_norm, score,
+    label)] with bbox normalized (x, y, w, h) in [0,1]. Self-disables on any
+    load/inference failure."""
+
+    def __init__(self, name, weights, conf=0.25, keep_labels=None):
+        self.name = name
+        self._spec = weights
+        self._conf = conf
+        self._keep = [k.lower() for k in keep_labels] if keep_labels else None
+        self._model = None
+        self._ok = True
+        self._lock = threading.Lock()
+
+    def _load(self):
+        if self._model is not None or not self._ok:
+            return
+        with self._lock:
+            if self._model is not None or not self._ok:
+                return
+            try:
+                from ultralytics import YOLO
+                path = _resolve(self._spec)
+                if path is None:
+                    self._ok = False
+                    return
+                self._model = YOLO(path)
+                # Disable ultralytics' load-time Conv+BN fusion. AutoBackend fuses
+                # the graph on the first predict; some checkpoints (yolo11n, the
+                # comic-panel model) crash that step with "'Conv' object has no
+                # attribute 'bn'" (a partially-fused / version-mismatched graph),
+                # which silently disabled those proposers (operator-flagged
+                # 2026-07-01). Unfused inference is correct — only marginally
+                # slower — and this is robust across ultralytics versions; if a
+                # future version ignores the override, the detect() guard below
+                # still self-disables the proposer instead of spamming per image.
+                inner = getattr(self._model, "model", None)
+                if inner is not None:
+                    inner.fuse = types.MethodType(lambda self, *a, **k: self, inner)
+                log.info("detector %s loaded (%s)", self.name, path)
+            except Exception as exc:   # noqa: BLE001
+                log.warning("detector %s disabled (load failed): %s", self.name, exc)
+                self._ok = False
+
+    def detect(self, image):
+        self._load()
+        if self._model is None:
+            return []
+        try:
+            res = self._model.predict(image, conf=self._conf, verbose=False)[0]
+        except Exception as exc:   # noqa: BLE001
+            # Permanently self-disable on the FIRST inference failure rather than
+            # re-throwing (and re-logging) on every image forever — an unfixable
+            # model fault degrades to "this proposer is off", logged once.
+            log.warning("detector %s disabled (inference failed): %s", self.name, exc)
+            self._ok = False
+            self._model = None
+            return []
+        iw, ih = image.size
+        names = getattr(res, "names", None) or {}
+        out = []
+        for b in res.boxes:
+            label = str(names.get(int(b.cls), int(b.cls))).lower()
+            if self._keep is not None and not any(k in label for k in self._keep):
+                continue
+            x0, y0, x1, y1 = (float(v) for v in b.xyxy[0].tolist())
+            out.append((
+                (x0 / iw, y0 / ih, (x1 - x0) / iw, (y1 - y0) / ih),
+                float(b.conf), label,
+            ))
+        return out
+
+
+class Proposers:
+    """The agent's proposer set, built from config. Each detector is optional —
+    an empty weight spec leaves that proposer off."""
+
+    def __init__(self, cfg):
+        self.cfg = cfg
+        self._person = (
+            YoloProposer("person-coco", cfg.person_weights,
+                         conf=cfg.person_conf, keep_labels=["person"])
+            if cfg.person_weights else None
+        )
+        self._anatomy = (
+            YoloProposer("anatomy", cfg.anatomy_weights, conf=cfg.anatomy_conf)
+            if cfg.anatomy_weights else None
+        )
+        self._panel = (
+            YoloProposer("panel", cfg.panel_weights, conf=cfg.panel_conf)
+            if cfg.panel_weights else None
+        )
+
+    def figures(self, image, base_boxes):
+        """Merge imgutils person boxes (base_boxes: [(bbox, score)]) with the
+        general COCO person detector → NMS'd figure boxes [(bbox, score, label)],
+        capped to the highest-scoring max_figures. Uncapped, a busy/huge image
+        (many characters) yields hundreds of boxes → hundreds of per-figure CCIP
+        calls + crops → a 30s+ job and an oversized submit (operator-flagged)."""
+        boxes = [(bb, sc if sc is not None else 1.0, "person") for bb, sc in base_boxes]
+        if self._person is not None:
+            boxes += self._person.detect(image)
+        return nms_merge(boxes)[: self.cfg.max_figures]   # nms_merge is score-desc
+
+    @staticmethod
+    def _top(detector, image, cap: int):
+        """Top-`cap` detections by score from an optional proposer (None → the
+        proposer is off → []). Shared by the anatomy + panel proposers, which
+        differ only in which detector and which cap."""
+        if detector is None:
+            return []
+        return sorted(detector.detect(image), key=lambda b: b[1], reverse=True)[:cap]
+
+    def components(self, image):
+        return self._top(self._anatomy, image, self.cfg.max_components)
+
+    def panels(self, image):
+        return self._top(self._panel, image, self.cfg.max_panels)

agent/fc_agent/embedder.py

@@ -58,12 +58,20 @@ class CropEmbedder:
 
     def embed(self, image: Image.Image) -> list[float]:
         """A crop → its embedding as a plain float list, ready to POST."""
+        return self.embed_batch([image])[0]
+
+    def embed_batch(self, images: list) -> list[list[float]]:
+        """Embed many crops in ONE forward pass — far better GPU utilisation +
+        only one lock acquisition than embedding each crop separately (which
+        starved the GPU and serialised the whole pool)."""
+        if not images:
+            return []
         self.load()
         torch = self._torch
-        enc = self._processor(images=image, return_tensors="pt")
+        enc = self._processor(images=images, return_tensors="pt")
         pixel_values = enc["pixel_values"].to(self._device, self._dt)
         with self._infer_lock, torch.no_grad():
             out = self._model.get_image_features(pixel_values=pixel_values)
         pooled = out.pooler_output if hasattr(out, "pooler_output") else out
-        vec = pooled[0].float().cpu().numpy().astype(np.float32).reshape(-1)
+        arr = pooled.float().cpu().numpy().astype(np.float32)
-        return vec.tolist()
+        return [row.reshape(-1).tolist() for row in arr]

agent/fc_agent/gpu.py

@@ -1,10 +1,24 @@
 """GPU load readout via nvidia-smi (present in the container thanks to the
 NVIDIA Container Toolkit's `utility` capability). Returns None if unavailable —
-the UI just shows n/a (e.g. CPU-fallback run)."""
+the UI just shows n/a (e.g. CPU-fallback run).
+
+Reads are CACHED and de-duplicated: the UI meter polls fast, /status reads it,
+and the autoscaler samples it — if each spawned its own `nvidia-smi` (slow on a
+busy GPU) those blocking subprocesses would pile up in the server's thread pool
+and make the Start/Stop buttons feel dead. So a short TTL serves recent callers
+from cache, and only ONE probe runs at a time (others get the last value)."""
 import subprocess
+import threading
+import time
+
+_TTL = 1.0                       # seconds a sample is reused before re-probing
+_lock = threading.Lock()
+_cache: dict | None = None
+_cache_t = 0.0
+_probing = False
 
 
-def read_gpu() -> dict | None:
+def _probe() -> dict | None:
     try:
         out = subprocess.run(
             [
@@ -28,3 +42,24 @@ def read_gpu() -> dict | None:
         }
     except (ValueError, IndexError):
         return None
+
+
+def read_gpu(max_age: float = _TTL) -> dict | None:
+    """Latest GPU reading, cached. Serves from cache when fresh; when stale,
+    exactly one caller re-probes while the rest get the last value — so request
+    threads never block behind more than one `nvidia-smi`."""
+    global _cache, _cache_t, _probing
+    now = time.monotonic()
+    with _lock:
+        fresh = _cache is not None and (now - _cache_t) < max_age
+        if fresh or _probing:        # fresh, or a probe is already running
+            return _cache
+        _probing = True
+    try:
+        val = _probe()
+    finally:
+        with _lock:
+            _cache = val
+            _cache_t = time.monotonic()
+            _probing = False
+    return val

agent/fc_agent/logbuf.py

@@ -0,0 +1,44 @@
+"""In-memory log ring buffer so the control UI can show recent agent logs
+(detector loads, job errors, autoscaler decisions, outage back-offs) without
+needing `docker logs`. A bounded deque holds the last N formatted lines; a
+logging.Handler appends to it; the UI polls /logs."""
+import logging
+from collections import deque
+
+LINES: deque[str] = deque(maxlen=400)
+
+
+class RingHandler(logging.Handler):
+    def emit(self, record: logging.LogRecord) -> None:
+        try:
+            LINES.append(self.format(record))
+        except Exception:
+            pass
+
+
+_installed = False
+
+
+def install(level: int = logging.INFO) -> None:
+    """Attach the ring handler to the root logger once. fc_agent module loggers
+    propagate to root, so their records land here."""
+    global _installed
+    if _installed:
+        return
+    _installed = True
+    h = RingHandler()
+    h.setFormatter(
+        logging.Formatter("%(asctime)s %(levelname)s %(name)s: %(message)s", "%H:%M:%S")
+    )
+    root = logging.getLogger()
+    root.addHandler(h)
+    if root.level == logging.NOTSET or root.level > level:
+        root.setLevel(level)
+    # Keep the buffer signal-rich: silence the chatty HTTP/download libs (every
+    # HF model fetch logs per-request) so the console shows agent activity —
+    # detector loads, job errors, autoscale moves — not request spam.
+    for noisy in (
+        "uvicorn.access", "ultralytics", "httpx", "httpcore",
+        "huggingface_hub", "urllib3", "filelock",
+    ):
+        logging.getLogger(noisy).setLevel(logging.WARNING)

agent/fc_agent/media.py

@@ -2,17 +2,77 @@
 (ffmpeg) at the cadence FC sends — so a video becomes a bag of per-frame
 instances, each with a timestamp."""
 import io
+import logging
 import os
+import signal
 import subprocess
 import tempfile
+import time
 
-from PIL import Image
+from PIL import Image, ImageFile
+
+from .throttle import PidReadMeter
+
+log = logging.getLogger("fc_agent.media")
+
+# Load slightly-truncated images (a few missing trailing bytes) instead of
+# raising — matches the server embedder. These are common in scraped libraries
+# and would otherwise fail the job 3× then error (operator-flagged 2026-06-30).
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+
+# Disable PIL's decompression-bomb guard: this is a TRUSTED local library, not an
+# untrusted upload surface, so a legitimately huge image (high-res scans/prints,
+# 90M+ pixels) must load. The default 89M-pixel limit only WARNS, but PIL raises
+# DecompressionBombError at 2× (~179M px) — which would fail those jobs outright
+# (operator-flagged 2026-06-30, images of 90–95M px).
+Image.MAX_IMAGE_PIXELS = None
 
 
 def is_video(mime: str) -> bool:
     return bool(mime) and (mime.startswith("video/") or mime in {"image/gif"})
 
 
+def _dhash(img: Image.Image, size: int = 8) -> int:
+    """Difference hash: compare adjacent pixels of a (size+1 × size) grayscale
+    thumbnail → a `size*size`-bit fingerprint. Cheap (64 comparisons on a 72-px
+    thumbnail) and robust to scaling/compression noise — near-identical frames
+    hash within a few bits, a real scene change moves many."""
+    small = img.convert("L").resize((size + 1, size))
+    px = list(small.getdata())
+    bits = 0
+    for row in range(size):
+        base = row * (size + 1)
+        for col in range(size):
+            bits = (bits << 1) | int(px[base + col] > px[base + col + 1])
+    return bits
+
+
+def dedupe_frames(
+    frames: list[tuple[float, Image.Image]], min_distance: int
+) -> list[tuple[float, Image.Image]]:
+    """Drop visually near-duplicate frames. A near-static video sampled into many
+    frames re-runs the WHOLE detect→CCIP→SigLIP chain on ~identical frames — the
+    dominant video load. Greedy perceptual-hash dedup: keep a frame only if its
+    dHash differs from every already-kept frame by >= min_distance bits (Hamming),
+    so a static run collapses to one frame while genuinely distinct scenes all
+    survive. Order + timestamps preserved. CPU-only (64-bit int XORs), so it runs
+    in the decode stage and spares the GPU the skipped frames entirely.
+
+    min_distance is the coarseness dial: higher keeps more frames (safer for brief
+    localized changes an 8×8 hash can miss), 0 disables. The first frame is always
+    kept (nothing to compare against)."""
+    if min_distance <= 0 or len(frames) <= 1:
+        return frames
+    kept: list[tuple[float, Image.Image]] = []
+    hashes: list[int] = []
+    for t, frame in frames:
+        h = _dhash(frame)
+        if all(bin(h ^ k).count("1") >= min_distance for k in hashes):
+            hashes.append(h)
+            kept.append((t, frame))
+    return kept
+
+
 def to_rgb(img: Image.Image) -> Image.Image:
     """RGB, flattening any transparency onto white first. A naive convert('RGB')
     on a palette-with-transparency image (common for character PNGs on a clear
@@ -32,32 +92,162 @@ def load_image(data: bytes) -> Image.Image:
     return to_rgb(Image.open(io.BytesIO(data)))
 
 
-def sample_frames(
+# ffmpeg reconnect flags — resume a dropped HTTP transfer (a slow/contended media
-    data: bytes, interval_seconds: float, max_frames: int
+# store can cut a long stream) instead of failing the whole job. Relies only on
+# HTTP + Range, which every FC deployment serves → environment-agnostic.
+_RECONNECT = [
+    "-reconnect", "1", "-reconnect_streamed", "1",
+    "-reconnect_on_network_error", "1", "-reconnect_delay_max", "5",
+]
+
+
+def _collect_frames(
+    tmp: str, interval: float, cap: int
 ) -> list[tuple[float, Image.Image]]:
-    """Extract up to max_frames frames at one-every-interval_seconds via ffmpeg.
+    out: list[tuple[float, Image.Image]] = []
-    Returns [(timestamp_seconds, frame)]. Empty on failure (caller falls back)."""
+    names = sorted(n for n in os.listdir(tmp) if n.startswith("f_"))
+    for i, name in enumerate(names[:cap]):
+        with Image.open(os.path.join(tmp, name)) as im:
+            out.append((round(i * interval, 2), to_rgb(im)))
+    return out
+
+
+def _terminate(proc: subprocess.Popen) -> None:
+    """Stop an ffmpeg cleanly, then hard-kill if it ignores SIGTERM."""
+    try:
+        # A bandwidth-paused (SIGSTOPped) process can't receive SIGTERM until it
+        # resumes — always CONT first so termination is prompt, not queued.
+        proc.send_signal(signal.SIGCONT)
+    except OSError:
+        pass
+    proc.terminate()
+    try:
+        proc.wait(timeout=2)
+    except subprocess.TimeoutExpired:
+        proc.kill()
+        try:
+            proc.wait(timeout=2)
+        except subprocess.TimeoutExpired:
+            pass
+
+
+def _pause(proc: subprocess.Popen, seconds: float, should_stop) -> bool:
+    """SIGSTOP ffmpeg for ~`seconds` of bandwidth debt, staying responsive to
+    Stop. While paused, the kernel socket buffer fills and TCP flow control
+    stalls curator's send side — that's the throttle. SIGCONT is ALWAYS sent
+    before returning. False = a Stop arrived mid-pause."""
+    try:
+        proc.send_signal(signal.SIGSTOP)
+    except OSError:
+        return True                     # already exited — nothing to pause
+    try:
+        end = time.monotonic() + seconds
+        while (left := end - time.monotonic()) > 0:
+            if should_stop and should_stop():
+                return False
+            time.sleep(min(0.5, left))
+        return True
+    finally:
+        try:
+            proc.send_signal(signal.SIGCONT)
+        except OSError:
+            pass
+
+
+def sample_frames_from_url(
+    url: str, interval_seconds: float, max_frames: int,
+    *, headers: str = "", timeout: float = 1200.0, should_stop=None,
+    governor=None,
+) -> tuple[list[tuple[float, Image.Image]], str | None]:
+    """Sample frames by pointing ffmpeg STRAIGHT at the media URL — it Range-reads
+    only the video index + up to max_frames worth of content, so the agent never
+    downloads the whole file (VR/4K originals run 800MB+ and would buffer ~1GB in
+    RAM and get cut off mid-download). Reconnect flags resume a dropped transfer;
+    the timeout is the per-video ceiling (a slow/reconnecting stream can otherwise
+    run for minutes). `should_stop` is polled while ffmpeg runs so a Stop KILLS the
+    subprocess at once — otherwise a downloader stuck in a long decode keeps the
+    agent "working" long after Stop. `governor` (the worker's shared TokenBucket)
+    meters ffmpeg's network reads from outside via /proc/<pid>/io and SIGSTOPs
+    the process into budget, so video streaming honors the same aggregate
+    bandwidth cap as still downloads.
+
+    Returns (frames, reason): frames is empty on failure/stop/timeout, and
+    `reason` then carries the SPECIFIC cause (ffmpeg's stderr tail / timeout) so
+    the caller can put it in the job's error — a bare "no frames" hid a filter
+    bug as "unprocessable" for weeks. None reason on success."""
     interval = max(0.5, float(interval_seconds or 4.0))
     cap = max(1, int(max_frames or 64))
+    hdr = ["-headers", headers] if headers else []
+    # select (NOT the fps filter): always keep the FIRST frame, then one per
+    # `interval` seconds of timestamp. fps=1/N emits round(duration/N) frames,
+    # which is ZERO for any clip shorter than ~N/2 seconds — a whole class of
+    # short animation loops failed as "unprocessable" that way (operator-flagged
+    # 2026-07-02: 0.5s/1.75s clips). scale=out_range=full converts limited-range
+    # yuv420p to full range so the mjpeg (jpg) encoder accepts it at default
+    # strictness instead of erroring on "non full-range YUV".
+    vf = (
+        f"select='isnan(prev_selected_t)+gte(t-prev_selected_t\\,{interval})',"
+        "scale=out_range=full"
+    )
     with tempfile.TemporaryDirectory() as tmp:
-        src = os.path.join(tmp, "in")
-        with open(src, "wb") as fh:
-            fh.write(data)
         pattern = os.path.join(tmp, "f_%05d.jpg")
+        cmd = ["ffmpeg", "-nostdin", "-loglevel", "error", *_RECONNECT, *hdr,
+               "-i", url, "-vf", vf, "-fps_mode", "vfr",
+               "-frames:v", str(cap), "-q:v", "3", pattern]
+        # ffmpeg's stderr goes to a file (not a PIPE, which could fill and
+        # deadlock; not DEVNULL, which is how a filter bug hid as "unprocessable"
+        # for weeks) — on failure its tail is logged so the operator can see WHY.
+        errpath = os.path.join(tmp, "stderr.txt")
         try:
-            subprocess.run(
+            with open(errpath, "wb") as errf:
-                [
+                proc = subprocess.Popen(
-                    "ffmpeg", "-nostdin", "-loglevel", "error", "-i", src,
+                    cmd, stdin=subprocess.DEVNULL,
-                    "-vf", f"fps=1/{interval}", "-frames:v", str(cap),
+                    stdout=subprocess.DEVNULL, stderr=errf,
-                    "-q:v", "3", pattern,
+                )
-                ],
+                meter = PidReadMeter(proc.pid) if governor is not None else None
-                check=True, timeout=600,
+                # Poll rather than block, so a Stop (or the per-video timeout) can
-            )
+                # kill a slow/wedged ffmpeg promptly instead of waiting it out.
-        except (subprocess.SubprocessError, FileNotFoundError):
+                start = time.monotonic()
-            return []
+                while True:
-        out: list[tuple[float, Image.Image]] = []
+                    try:
-        names = sorted(n for n in os.listdir(tmp) if n.startswith("f_"))
+                        proc.wait(timeout=0.5)
-        for i, name in enumerate(names[:cap]):
+                        break
-            with Image.open(os.path.join(tmp, name)) as im:
+                    except subprocess.TimeoutExpired:
-                out.append((round(i * interval, 2), to_rgb(im)))
+                        stopped = should_stop and should_stop()
-        return out
+                        if stopped or (time.monotonic() - start > timeout):
+                            _terminate(proc)
+                            if stopped:
+                                return [], "stopped"
+                            log.warning("ffmpeg timed out after %.0fs: %s",
+                                        timeout, url)
+                            return [], f"ffmpeg timed out after {timeout:.0f}s"
+                        if meter is not None:
+                            read = meter.delta()
+                            if read is None:      # /proc gone → stop governing
+                                meter = None
+                            elif (debt := governor.charge(read)) > 0:
+                                # Over budget: pause ffmpeg until the bucket
+                                # recovers. Pause time counts toward `timeout`
+                                # (it stays the wedge backstop either way).
+                                if not _pause(proc, debt, should_stop):
+                                    _terminate(proc)
+                                    return [], "stopped"
+        except (OSError, ValueError) as exc:
+            return [], f"ffmpeg not runnable: {exc}"
+        frames = _collect_frames(tmp, interval, cap)
+        if not frames:
+            reason = f"ffmpeg exit {proc.returncode}: {_tail(errpath)}"
+            log.warning("ffmpeg produced no frames for %s — %s", url, reason)
+            return [], reason
+        return frames, None
+
+
+def _tail(path: str, limit: int = 300) -> str:
+    """Last `limit` chars of a (stderr) file, flattened — for failure logs."""
+    try:
+        with open(path, "rb") as f:
+            f.seek(0, os.SEEK_END)
+            f.seek(max(0, f.tell() - limit))
+            return f.read().decode("utf-8", "replace").replace("\n", " ").strip()
+    except OSError:
+        return "?"

agent/fc_agent/throttle.py

@@ -0,0 +1,111 @@
+"""Global download-bandwidth governor (one token bucket for the whole agent).
+
+The agent lives on someone's desktop and shares that desktop's network —
+typically WiFi, where saturating the link doesn't just slow other apps: it
+bufferbloats the airtime (RTT 21→45ms) and collapses EVERY connection,
+the operator's browser included. Measured 2026-07-02: the idle link moved
+~38 MB/s single-stream, but under the 8-downloader sweep every stream on the
+machine crawled at ~1-1.5 MB/s. So the cap is on the AGGREGATE, not per
+stream: still downloads pump their chunks through take(), and ffmpeg video
+streams — whose sockets live in a subprocess we can't wrap — are metered from
+outside via /proc/<pid>/io and paused (SIGSTOP) into budget using charge()'s
+debt signal; TCP flow control then stalls the sender while ffmpeg sleeps.
+
+Accounting is post-paid (charge the bytes first, then wait out any debt): the
+bytes have already crossed the network by the time we count them, and it means
+a chunk larger than one second of budget can never deadlock the bucket.
+Stdlib-only on purpose — unit-tested in CI, where the agent's ML deps
+don't exist.
+"""
+import threading
+import time
+
+
+class TokenBucket:
+    """Thread-safe token bucket in bytes/second. rate 0 = unlimited.
+
+    `consumed` is the monotonic total of bytes charged (throttled or not) —
+    the worker's rate loop derives the UI's "net MB/s" readout from it.
+    """
+
+    def __init__(self, rate_bytes_per_s: float = 0.0):
+        self._cond = threading.Condition()
+        self._rate = max(0.0, float(rate_bytes_per_s))
+        # Burst = one second of budget: enough that chunked reads stay smooth,
+        # small enough that a burst can't meaningfully lift the average.
+        self._level = self._rate
+        self._stamp = time.monotonic()
+        self.consumed = 0
+
+    @property
+    def rate(self) -> float:
+        return self._rate
+
+    def set_rate(self, rate_bytes_per_s: float) -> None:
+        """Retune live (the UI dial). Waiters re-check immediately, so raising
+        the cap (or lifting it with 0) unblocks a mid-download wait at once."""
+        with self._cond:
+            self._refill_locked()          # settle elapsed time at the OLD rate
+            self._rate = max(0.0, float(rate_bytes_per_s))
+            self._level = min(self._level, self._rate)
+            self._cond.notify_all()
+
+    def _refill_locked(self) -> None:
+        now = time.monotonic()
+        self._level = min(self._rate, self._level + (now - self._stamp) * self._rate)
+        self._stamp = now
+
+    def take(self, n: int) -> None:
+        """Charge n bytes and block until the budget recovers (stills path)."""
+        with self._cond:
+            self.consumed += n
+            if self._rate <= 0:
+                return
+            self._refill_locked()
+            self._level -= n
+            while self._level < 0:
+                # Wake early on set_rate; cap the wait so a big debt is paid in
+                # re-checked slices rather than one long uninterruptible sleep.
+                self._cond.wait(min(-self._level / self._rate, 0.5))
+                if self._rate <= 0:
+                    return
+                self._refill_locked()
+
+    def charge(self, n: int) -> float:
+        """Charge n bytes WITHOUT blocking; return seconds of debt (0 = within
+        budget). The ffmpeg governor can't block the subprocess's own reads, so
+        it SIGSTOPs the process for (about) the returned debt instead."""
+        with self._cond:
+            self.consumed += n
+            if self._rate <= 0:
+                return 0.0
+            self._refill_locked()
+            self._level -= n
+            return max(0.0, -self._level / self._rate)
+
+
+class PidReadMeter:
+    """Cumulative read-bytes meter for a subprocess, via /proc/<pid>/io.
+
+    `rchar` counts every read() syscall's bytes — for a streaming ffmpeg the
+    network reads dominate, so the delta is a good-enough aggregate-bandwidth
+    signal (it's a governor, not a billing meter). Returns None when /proc is
+    unavailable (process exited, or a non-Linux host): the caller then simply
+    doesn't govern — degrade to unthrottled rather than break video sampling.
+    """
+
+    def __init__(self, pid: int):
+        self._path = f"/proc/{pid}/io"
+        self._last = 0
+
+    def delta(self) -> int | None:
+        try:
+            with open(self._path, "rb") as f:
+                for line in f:
+                    if line.startswith(b"rchar:"):
+                        total = int(line.split()[1])
+                        d, self._last = total - self._last, total
+                        return max(0, d)
+        except (OSError, ValueError):
+            return None
+        return None

agent/requirements.txt

@@ -7,6 +7,9 @@ onnxruntime-gpu
 # Dockerfile from the CUDA-12.4 wheel index so the GPU build is deterministic;
 # transformers loads whatever SigLIP-family model the server announces.
 transformers>=4.45
+# Crop PROPOSERS — small YOLO detectors (booru_yolo anatomy, COCO person, comic
+# panel) that decide where to crop. Uses the torch already installed above.
+ultralytics>=8.3
 # Control surface + HTTP.
 fastapi
 uvicorn[standard]

agent/ruff.toml

@@ -0,0 +1,7 @@
+# The agent runs on the CUDA base image's Python 3.12 (Ubuntu 24.04) — NOT the
+# 3.14 that CI's ci-python image and the repo-root ruff.toml target. Pin the
+# agent to py312 so ruff enforces 3.12 compatibility and never auto-applies a
+# 3.14-only fix (e.g. unquoting a self-referential annotation, which PEP 649
+# makes safe on 3.14 but NameErrors on 3.12). Inherit the root lint rules.
+extend = "../ruff.toml"
+target-version = "py312"

alembic/versions/0065_embedder_model_name.py

@@ -0,0 +1,35 @@
+"""ml_settings: embedder_model_name (#1190 operator model swap)
+
+The embedder MODEL VERSION was already a setting (and stamps image_record.
+siglip_model_version); the HF model NAME was env-only, so an operator couldn't
+actually point the pipeline at a different embedder. Storing the name as a
+setting makes the model an operator choice: set name + version → re-embed (the
+GPU agent) → retrain heads. Default = the current SigLIP so400m.
+
+Revision ID: 0065
+Revises: 0064
+Create Date: 2026-06-30
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0065"
+down_revision: Union[str, None] = "0064"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "ml_settings",
+        sa.Column(
+            "embedder_model_name", sa.String(length=128), nullable=False,
+            server_default="google/siglip-so400m-patch14-384",
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("ml_settings", "embedder_model_name")

alembic/versions/0066_drop_centroids.py

@@ -0,0 +1,57 @@
+"""drop the dead per-tag centroid subsystem (#1189 cleanup)
+
+The v2 pivot replaced per-tag SigLIP centroids with learned heads + CCIP.
+Nothing read the centroids anymore — they were recomputed (on merge + a daily
+beat) but never consumed for suggestions or auto-apply. Remove the storage +
+its two now-unused settings columns. (The recompute tasks, beat, endpoint,
+service, and UI card are removed in the same change.)
+
+Revision ID: 0066
+Revises: 0065
+Create Date: 2026-06-30
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0066"
+down_revision: Union[str, None] = "0065"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.drop_table("tag_reference_embedding")
+    op.drop_column("ml_settings", "centroid_similarity_threshold")
+    op.drop_column("ml_settings", "min_reference_images")
+
+
+def downgrade() -> None:
+    op.add_column(
+        "ml_settings",
+        sa.Column(
+            "min_reference_images", sa.Integer(), nullable=False,
+            server_default="5",
+        ),
+    )
+    op.add_column(
+        "ml_settings",
+        sa.Column(
+            "centroid_similarity_threshold", sa.Float(), nullable=False,
+            server_default="0.55",
+        ),
+    )
+    op.create_table(
+        "tag_reference_embedding",
+        sa.Column("tag_id", sa.Integer(), nullable=False),
+        sa.Column("embedding", sa.LargeBinary(), nullable=False),
+        sa.Column("reference_count", sa.Integer(), nullable=False),
+        sa.Column("model_version", sa.String(length=128), nullable=False),
+        sa.Column(
+            "updated_at", sa.DateTime(timezone=True),
+            server_default=sa.func.now(), nullable=False,
+        ),
+        sa.ForeignKeyConstraint(["tag_id"], ["tag.id"], ondelete="CASCADE"),
+        sa.PrimaryKeyConstraint("tag_id"),
+    )

alembic/versions/0067_retire_camie_allowlist.py

@@ -0,0 +1,66 @@
+"""retire the Camie tagger + allowlist bulk-apply (#1189)
+
+The v2 pivot made heads + CCIP the tag source and head auto-apply the earned
+propagation. The Camie tagger ran only to feed the allowlist bulk-apply (its
+predictions had no other consumer), and the allowlist was a second, un-earned
+auto-apply path parallel to heads. Both are retired — drop their storage.
+
+(image_prediction = Camie's per-image predictions; tag_allowlist = the bulk-
+apply allowlist. Nothing references INTO these tables, so the drop is clean.)
+
+Revision ID: 0067
+Revises: 0066
+Create Date: 2026-06-30
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0067"
+down_revision: Union[str, None] = "0066"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.drop_table("image_prediction")
+    op.drop_table("tag_allowlist")
+
+
+def downgrade() -> None:
+    op.create_table(
+        "tag_allowlist",
+        sa.Column("tag_id", sa.Integer(), nullable=False),
+        sa.Column(
+            "min_confidence", sa.Float(), nullable=False, server_default="0.9"
+        ),
+        sa.Column(
+            "created_at", sa.DateTime(timezone=True),
+            server_default=sa.func.now(), nullable=False,
+        ),
+        sa.ForeignKeyConstraint(["tag_id"], ["tag.id"], ondelete="CASCADE"),
+        sa.PrimaryKeyConstraint("tag_id"),
+        sa.CheckConstraint(
+            "min_confidence >= 0 AND min_confidence <= 1",
+            name="ck_tag_allowlist_confidence_range",
+        ),
+    )
+    op.create_table(
+        "image_prediction",
+        sa.Column("id", sa.Integer(), primary_key=True),
+        sa.Column("image_record_id", sa.Integer(), nullable=False),
+        sa.Column("raw_name", sa.String(length=255), nullable=False),
+        sa.Column("category", sa.String(length=32), nullable=False),
+        sa.Column("score", sa.Float(), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["image_record_id"], ["image_record.id"], ondelete="CASCADE"
+        ),
+    )
+    op.create_index(
+        "ix_image_prediction_image", "image_prediction", ["image_record_id"]
+    )
+    op.create_index(
+        "ix_image_prediction_name_score", "image_prediction",
+        ["raw_name", "score"],
+    )

alembic/versions/0068_drop_dead_tagger_settings.py

@@ -0,0 +1,80 @@
+"""drop dead tagger/suggestion settings + columns left after Camie retirement (#1199)
+
+Hygiene follow-up to #1189. These were left inert to bound that change; nothing
+reads them now:
+- ml_settings: tagger_store_floor + tagger_model_version (only the deleted Camie
+  tagger used them), suggestion_threshold_character/general (already dead pre-
+  retirement — scoring uses per-head thresholds), video_min_tag_frames (only the
+  deleted video-prediction aggregator used it).
+- image_record: tagger_model_version (no writer now), centroid_scores (long-dead
+  JSON cache, no reader).
+
+Revision ID: 0068
+Revises: 0067
+Create Date: 2026-06-30
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0068"
+down_revision: Union[str, None] = "0067"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.drop_column("ml_settings", "suggestion_threshold_character")
+    op.drop_column("ml_settings", "suggestion_threshold_general")
+    op.drop_column("ml_settings", "tagger_store_floor")
+    op.drop_column("ml_settings", "video_min_tag_frames")
+    op.drop_column("ml_settings", "tagger_model_version")
+    op.drop_column("image_record", "tagger_model_version")
+    op.drop_column("image_record", "centroid_scores")
+
+
+def downgrade() -> None:
+    op.add_column(
+        "image_record",
+        sa.Column("centroid_scores", sa.JSON(), nullable=True),
+    )
+    op.add_column(
+        "image_record",
+        sa.Column("tagger_model_version", sa.String(length=128), nullable=True),
+    )
+    op.add_column(
+        "ml_settings",
+        sa.Column(
+            "tagger_model_version", sa.String(length=128), nullable=False,
+            server_default="camie-tagger-v2",
+        ),
+    )
+    op.add_column(
+        "ml_settings",
+        sa.Column(
+            "video_min_tag_frames", sa.Integer(), nullable=False,
+            server_default="3",
+        ),
+    )
+    op.add_column(
+        "ml_settings",
+        sa.Column(
+            "tagger_store_floor", sa.Float(), nullable=False,
+            server_default="0.7",
+        ),
+    )
+    op.add_column(
+        "ml_settings",
+        sa.Column(
+            "suggestion_threshold_general", sa.Float(), nullable=False,
+            server_default="0.7",
+        ),
+    )
+    op.add_column(
+        "ml_settings",
+        sa.Column(
+            "suggestion_threshold_character", sa.Float(), nullable=False,
+            server_default="0.7",
+        ),
+    )

alembic/versions/0069_default_siglip2.py

@@ -0,0 +1,51 @@
+"""default the embedder to SigLIP 2 — for FRESH installs only (#1203)
+
+Make SigLIP 2 (so400m, 512px; a 1152-d drop-in) the default embedder. New
+installs start on it. An EXISTING library is NOT touched: flipping its stored
+embedder version would mark every embedding stale (the scorer is version-gated)
+and kill suggestions until a full re-embed+retrain — so an existing instance
+switches deliberately via Settings → GPU agent → Embedding model → Re-embed →
+Retrain. We detect "fresh" by the absence of any embedded image.
+
+Revision ID: 0069
+Revises: 0068
+Create Date: 2026-06-30
+"""
+from typing import Sequence, Union
+
+from alembic import op
+
+revision: str = "0069"
+down_revision: Union[str, None] = "0068"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+_NEW_NAME = "google/siglip2-so400m-patch16-512"
+_NEW_VERSION = "siglip2-so400m-patch16-512"
+_OLD_NAME = "google/siglip-so400m-patch14-384"
+_OLD_VERSION = "siglip-so400m-patch14-384"
+
+
+def upgrade() -> None:
+    # Fresh install (nothing embedded yet) → adopt SigLIP 2.
+    op.execute(
+        f"""
+        UPDATE ml_settings SET
+            embedder_model_name = '{_NEW_NAME}',
+            embedder_model_version = '{_NEW_VERSION}'
+        WHERE NOT EXISTS (
+            SELECT 1 FROM image_record WHERE siglip_embedding IS NOT NULL
+        )
+        """
+    )
+    op.alter_column("ml_settings", "embedder_model_name", server_default=_NEW_NAME)
+    op.alter_column(
+        "ml_settings", "embedder_model_version", server_default=_NEW_VERSION
+    )
+
+
+def downgrade() -> None:
+    op.alter_column("ml_settings", "embedder_model_name", server_default=_OLD_NAME)
+    op.alter_column(
+        "ml_settings", "embedder_model_version", server_default=_OLD_VERSION
+    )

alembic/versions/0070_gpu_job_lease_indexes.py

@@ -0,0 +1,44 @@
+"""partial indexes so GPU-job leasing stays O(batch), not O(completed)
+
+The lease claims the lowest-id pending (or expired-leased) jobs. With only a
+plain `status` index, `... ORDER BY id LIMIT n` walked the primary-key index from
+the start, skipping the entire prefix of already-done/error rows before reaching
+pending ones — so leasing slowed to a crawl as `done` piled up (the whole reason
+throughput fell off a cliff mid-run and /status stalled). Two partial indexes fix
+it: the pending one is id-ordered so the hot path reads just the first n entries,
+and the leased-expiry one keeps the crash-recovery reclaim + the orphan sweep
+cheap. They cover only the small live slice of the table, so they stay tiny even
+as the done/error history grows to millions.
+
+Revision ID: 0070
+Revises: 0069
+Create Date: 2026-06-30
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0070"
+down_revision: Union[str, None] = "0069"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Hot path: lowest-id pending jobs. Index on id, restricted to pending, so
+    # `WHERE status='pending' ORDER BY id LIMIT n` is a short index-order scan.
+    op.create_index(
+        "ix_gpu_job_pending", "gpu_job", ["id"],
+        postgresql_where=sa.text("status = 'pending'"),
+    )
+    # Crash-recovery: expired leases, for the lease backstop + recover_orphaned.
+    op.create_index(
+        "ix_gpu_job_leased_expires", "gpu_job", ["lease_expires_at"],
+        postgresql_where=sa.text("status = 'leased'"),
+    )
+
+
+def downgrade() -> None:
+    op.drop_index("ix_gpu_job_leased_expires", table_name="gpu_job")
+    op.drop_index("ix_gpu_job_pending", table_name="gpu_job")

alembic/versions/0071_image_record_earliest_post_date.py

@@ -0,0 +1,80 @@
+"""image_record.earliest_post_date: original-publish gallery sort key + index
+
+Revision ID: 0071
+Revises: 0070
+Create Date: 2026-07-01
+
+effective_date (0035) keys off the PRIMARY post — which is often the repost /
+download the file actually came from — and falls back to created_at, so the
+gallery's default order surfaces download dates rather than when content was
+first posted (operator-flagged 2026-07-01). Materialize a second sort key,
+earliest_post_date = MIN(post_date) across ALL of an image's provenance posts
+(every post it appears in), falling back to created_at only when no linked post
+carries a date. Indexed (DESC, id DESC) so the "post date" gallery sort is an
+index range scan just like effective_date.
+
+Backfill mirrors 0035: created_at baseline, then override with the MIN over
+image_provenance ⋈ post. New rows get the created_at-equivalent server default;
+services/importer.py recomputes it whenever a dated post is linked.
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0071"
+down_revision: Union[str, None] = "0070"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Add nullable first so the backfill can populate before NOT NULL.
+    op.add_column(
+        "image_record",
+        sa.Column("earliest_post_date", sa.DateTime(timezone=True), nullable=True),
+    )
+    # Baseline: download date. Set-based (no per-row binds) → immune to the
+    # 65535 bind-parameter ceiling regardless of library size.
+    op.execute(
+        """
+        UPDATE image_record
+        SET earliest_post_date = created_at
+        """
+    )
+    # Override with the earliest post_date across EVERY post the image appears
+    # in (image_provenance is the many-to-many edge; ignore posts with no date).
+    op.execute(
+        """
+        UPDATE image_record AS ir
+        SET earliest_post_date = sub.min_date
+        FROM (
+            SELECT ip.image_record_id AS iid, MIN(p.post_date) AS min_date
+            FROM image_provenance AS ip
+            JOIN post AS p ON p.id = ip.post_id
+            WHERE p.post_date IS NOT NULL
+            GROUP BY ip.image_record_id
+        ) AS sub
+        WHERE ir.id = sub.iid
+        """
+    )
+    op.alter_column(
+        "image_record",
+        "earliest_post_date",
+        nullable=False,
+        server_default=sa.text("now()"),
+    )
+    # DESC/DESC matches the gallery's ORDER BY earliest_post_date DESC, id DESC
+    # so the "post date" scroll is a forward index scan; raw SQL because
+    # alembic's column list doesn't express per-column DESC cleanly.
+    op.execute(
+        "CREATE INDEX ix_image_record_earliest_post_date "
+        "ON image_record (earliest_post_date DESC, id DESC)"
+    )
+
+
+def downgrade() -> None:
+    op.drop_index(
+        "ix_image_record_earliest_post_date", table_name="image_record"
+    )
+    op.drop_column("image_record", "earliest_post_date")

alembic/versions/0072_gpu_job_triage_status.py

@@ -0,0 +1,32 @@
+"""gpu_job.triage_status — the probe's verdict on an errored job's FILE
+
+Failure triage (#125): a periodic sweep probes each errored image's file
+(sha256 + decode, verify_integrity's machinery) exactly once and stores the
+verdict here — 'defect' (the file is bad: recovery material, excluded from
+/retry_errors) or 'file_ok' (failure was operational, safe to retry). NULL
+means not yet probed; selecting on NULL is what makes the sweep resumable.
+No index: the errored slice the sweep scans is tiny by design (tombstones).
+
+Revision ID: 0072
+Revises: 0071
+Create Date: 2026-07-02
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0072"
+down_revision: Union[str, None] = "0071"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "gpu_job", sa.Column("triage_status", sa.String(16), nullable=True)
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("gpu_job", "triage_status")

alembic/versions/0073_drop_tag_eval_run.py

@@ -0,0 +1,46 @@
+"""drop tag_eval_run — the head-vs-centroid eval harness is retired
+
+The eval (#1130) existed to prove the heads tagging spine on the operator's own
+data. It did; the operator accepted the system and retired the harness
+(2026-07-02) — card, API, task, model and this table all go. The eval's data
+loaders + metric helpers live on in services/ml/training_data.py, where the
+production heads trainer uses them nightly.
+
+Revision ID: 0073
+Revises: 0072
+Create Date: 2026-07-02
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+revision: str = "0073"
+down_revision: Union[str, None] = "0072"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.drop_index("ix_tag_eval_run_status", table_name="tag_eval_run")
+    op.drop_table("tag_eval_run")
+
+
+def downgrade() -> None:
+    # Recreates the shape from 0056 (data is not restorable).
+    op.create_table(
+        "tag_eval_run",
+        sa.Column("id", sa.Integer(), primary_key=True),
+        sa.Column("params", postgresql.JSONB(), nullable=False),
+        sa.Column("status", sa.String(length=16), nullable=False,
+                  server_default="running"),
+        sa.Column("started_at", sa.DateTime(timezone=True), nullable=False,
+                  server_default=sa.func.now()),
+        sa.Column("finished_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column("report", postgresql.JSONB(), nullable=True),
+        sa.Column("error", sa.Text(), nullable=True),
+        sa.Column("last_progress_at", sa.DateTime(timezone=True),
+                  nullable=True),
+    )
+    op.create_index("ix_tag_eval_run_status", "tag_eval_run", ["status"])

alembic/versions/0074_ml_settings_cpu_embed_enabled.py

@@ -0,0 +1,35 @@
+"""ml_settings.cpu_embed_enabled — the CPU embed fallback becomes a switch
+
+B3 (operator 2026-07-02): the ml-worker's only processing role is the CPU
+whole-image embed for stacks without a GPU agent. ON by default (a fresh
+install works agent-less); agent-equipped stacks that drop the ml-worker
+container turn it off so import hooks stop queueing embed work into a queue
+nothing consumes — the daily GPU 'embed' backfill covers those images.
+
+Revision ID: 0074
+Revises: 0073
+Create Date: 2026-07-02
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0074"
+down_revision: Union[str, None] = "0073"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "ml_settings",
+        sa.Column(
+            "cpu_embed_enabled", sa.Boolean(), nullable=False,
+            server_default=sa.true(),
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("ml_settings", "cpu_embed_enabled")

backend/app/__init__.py

@@ -34,6 +34,23 @@ def create_app() -> Quart:
     app = Quart(__name__)
     app.secret_key = cfg.secret_key
 
+    # Stream files in 4 MiB chunks instead of Quart's 8 KiB default. The image
+    # library lives on a CIFS/SMB share (mounted rsize=4 MiB), so 8 KiB reads
+    # meant ~19k network round-trips for one large original — 30–58s downloads
+    # that starved both the GPU agent and the browser (operator-flagged
+    # 2026-07-01). 4 MiB matches the mount's read size → one round-trip per read,
+    # ~500× fewer. buffer_size is the MAX read, so small thumbnails still read in
+    # a single gulp, and Range/mime/ETag/conditional handling lives on Response,
+    # so this keeps all of it. Guarded so a future Quart-internal change can't
+    # break boot — worst case we fall back to the slow default.
+    try:
+        from quart.wrappers.response import FileBody
+        FileBody.buffer_size = 4 * 1024 * 1024
+    except Exception:
+        logging.getLogger(__name__).warning(
+            "could not raise FileBody.buffer_size — file serving stays on 8 KiB chunks"
+        )
+
     for bp in all_blueprints():
         app.register_blueprint(bp)
     # Registered last so /api/* routes win over the SPA catch-all.

backend/app/api/__init__.py

@@ -16,7 +16,6 @@ api_bp.add_url_rule("/health", view_func=health.get_health, methods=["GET"])
 def all_blueprints() -> list[Blueprint]:
     from .admin import admin_bp
     from .aliases import aliases_bp
-    from .allowlist import allowlist_bp
     from .artist import artist_bp
     from .artists import artists_bp
     from .attachments import attachments_bp
@@ -39,7 +38,6 @@ def all_blueprints() -> list[Blueprint]:
     from .suggestions import suggestions_bp
     from .system_activity import system_activity_bp
     from .system_backup import system_backup_bp
-    from .tag_eval import tag_eval_bp
     from .tags import tags_bp
     from .thumbnails import thumbnails_bp
     return [
@@ -58,9 +56,7 @@ def all_blueprints() -> list[Blueprint]:
         cleanup_bp,
         import_admin_bp,
         suggestions_bp,
-        allowlist_bp,
         aliases_bp,
-        tag_eval_bp,
         heads_bp,
         gpu_bp,
         ccip_bp,

backend/app/api/admin.py

@@ -1,13 +1,12 @@
 """FC-3k: /api/admin — destructive admin actions.
 
-Five action surfaces:
+Action surfaces:
   POST   /api/admin/artists/<slug>/cascade-delete  (Tier C)
   POST   /api/admin/images/bulk-delete             (Tier C)
   DELETE /api/admin/tags/<int:tag_id>              (Tier B)
   POST   /api/admin/tags/<int:dest_id>/merge       (Tier B)
   POST   /api/admin/tags/prune-unused              (Tier A)
   POST   /api/admin/posts/prune-bare               (Tier A)
-  POST   /api/admin/tags/purge-legacy              (Tier A)
   GET    /api/admin/tags/<int:tag_id>/usage-count  (helper)
 
 Tier-C ops take a dry_run body flag (returns projection inline,
@@ -277,31 +276,48 @@ async def posts_reconcile_duplicates():
     return await _run_dry_run_op(reconcile_duplicate_posts, source_id=source_id)
 
 
-@admin_bp.route("/tags/purge-legacy", methods=["POST"])
+def _reset_content_confirm_token(projection: dict) -> str:
-async def tags_purge_legacy():
+    """Stable 8-hex token derived from the live counts (mirrors the Tier-C
-    """Tier-A: delete legacy IR-migration tags — archive/post/artist
+    bulk-delete token): it changes whenever the data changes, so the apply can
-    kinds (e.g. `BlenderKnight:Hannah_BJ_Loops`) PLUS general tags with
+    only ever run against numbers the operator just previewed."""
-    a legacy name prefix (`source:*`, from IR's source kind that fell
+    canon = f"reset-content:{projection.get('count')}:{projection.get('applications')}"
-    back to general). dry-run preview returns per-kind + per-prefix
+    return hashlib.sha256(canon.encode("utf-8")).hexdigest()[:8]
-    counts + a sample so the UI shows exactly what'll go before the
-    operator confirms with dry_run=false."""
-    from ..services.cleanup_service import purge_legacy_tags
-
-    return await _run_dry_run_op(purge_legacy_tags)
 
 
 @admin_bp.route("/tags/reset-content", methods=["POST"])
 async def tags_reset_content():
-    """Tier-A: delete ALL general + character tags (the Camie-suggestable
+    """Full-instance reset of the CONTENT vocabulary: deletes ALL general +
-    content vocabulary) so the operator can re-tag from scratch via
+    character tags and their image applications — INCLUDING the examples the
-    auto-suggest. fandom + series tags + series_page ordering are preserved,
+    tagging heads learned from. Suggestions do NOT repopulate on their own
-    and image_prediction rows are untouched so suggestions repopulate.
+    (the Camie predictions that once did are long retired): the operator
-    dry-run preview returns per-kind counts + applications + a sample so the
+    re-tags from scratch and the heads retrain from the new signal. fandom +
-    UI shows exactly what'll go before the operator confirms (dry_run=false).
+    series tags + series_page ordering are preserved.
-    Irreversible except via DB backup restore."""
+
+    Deliberately Tier-C-gated despite the Tier-A shape (operator 2026-07-02:
+    the full reset stays, but behind extra steps): dry_run returns the
+    projection + a `confirm` token derived from the live counts; the apply
+    must echo that token back or it is rejected."""
     from ..services.cleanup_service import reset_content_tagging
 
-    return await _run_dry_run_op(reset_content_tagging)
+    body = await request.get_json(silent=True) or {}
+    dry_run = bool(body.get("dry_run", False))
+    async with get_session() as session:
+        projection = await session.run_sync(
+            lambda s: reset_content_tagging(s, dry_run=True)
+        )
+        token = _reset_content_confirm_token(projection)
+        if dry_run:
+            projection["confirm"] = token
+            return jsonify(projection)
+        if str(body.get("confirm", "")) != token:
+            return _bad(
+                "confirm_mismatch",
+                detail="run a fresh preview and echo its confirm token",
+            )
+        result = await session.run_sync(
+            lambda s: reset_content_tagging(s, dry_run=False)
+        )
+    return jsonify(result)
 
 
 @admin_bp.route("/tags/normalize", methods=["POST"])

backend/app/api/allowlist.py

@@ -1,84 +0,0 @@
-"""Allowlist API: list, adjust threshold, remove."""
-
-from quart import Blueprint, jsonify, request
-
-from ..extensions import get_session
-from ..models import TagAllowlist
-from ..services.ml.allowlist import AllowlistService
-
-allowlist_bp = Blueprint("allowlist", __name__, url_prefix="/api")
-
-
-@allowlist_bp.route("/allowlist", methods=["GET"])
-async def list_allowlist():
-    async with get_session() as session:
-        rows = await AllowlistService(session).list_all()
-    return jsonify(
-        [
-            {
-                "tag_id": r.tag_id,
-                "tag_name": r.tag_name,
-                "tag_kind": r.tag_kind,
-                "min_confidence": r.min_confidence,
-                "applied_count": r.applied_count,
-                "coverage_count": r.coverage_count,
-            }
-            for r in rows
-        ]
-    )
-
-
-@allowlist_bp.route("/tags/<int:tag_id>/allowlist/coverage", methods=["GET"])
-async def coverage(tag_id: int):
-    """Live "at threshold T, a sweep would cover ~N images" projection for the
-    allowlist tuning dashboard. Defaults to the tag's stored threshold."""
-    raw = request.args.get("threshold")
-    async with get_session() as session:
-        svc = AllowlistService(session)
-        if raw is not None:
-            try:
-                threshold = float(raw)
-            except ValueError:
-                return jsonify({"error": "threshold must be a float"}), 400
-            if not (0 < threshold <= 1):
-                return jsonify({"error": "threshold must be in (0, 1]"}), 400
-        else:
-            row = await session.get(TagAllowlist, tag_id)
-            if row is None:
-                return jsonify({"error": "not on allowlist"}), 404
-            threshold = row.min_confidence
-        count = await svc.coverage(tag_id, threshold)
-    return jsonify({"count": count, "threshold": threshold})
-
-
-@allowlist_bp.route("/tags/<int:tag_id>/allowlist", methods=["GET"])
-async def get_one(tag_id: int):
-    async with get_session() as session:
-        row = await session.get(TagAllowlist, tag_id)
-    if row is None:
-        return jsonify({"error": "not on allowlist"}), 404
-    return jsonify(
-        {"min_confidence": row.min_confidence, "added_at": row.added_at.isoformat()}
-    )
-
-
-@allowlist_bp.route("/tags/<int:tag_id>/allowlist", methods=["PATCH"])
-async def patch_threshold(tag_id: int):
-    body = await request.get_json()
-    if not body or "min_confidence" not in body:
-        return jsonify({"error": "min_confidence required"}), 400
-    mc = float(body["min_confidence"])
-    if not (0 < mc <= 1):
-        return jsonify({"error": "min_confidence must be in (0, 1]"}), 400
-    async with get_session() as session:
-        await AllowlistService(session).update_threshold(tag_id, mc)
-        await session.commit()
-    return "", 204
-
-
-@allowlist_bp.route("/tags/<int:tag_id>/allowlist", methods=["DELETE"])
-async def remove(tag_id: int):
-    async with get_session() as session:
-        await AllowlistService(session).remove(tag_id)
-        await session.commit()
-    return "", 204

backend/app/api/gallery.py

@@ -44,7 +44,8 @@ def _parse_filters():
       the image must match at least one tag from EACH group (groups ANDed).
     - `tag_not` is a comma-separated exclude list (image must carry none).
 
-    `media` is image|video; `sort` is newest|oldest; `platform` selects one
+    `media` is image|video; `sort` is newest|oldest|posted_new|posted_old
+    (default posted_new); `platform` selects one
     platform (or the UNSOURCED_PLATFORM sentinel); `untagged`/`no_artist` are
     boolean flags; `date_from`/`date_to` are inclusive calendar-day bounds
     (date_to is widened by a day so the whole day is covered by the service's
@@ -67,8 +68,12 @@ def _parse_filters():
     artist_id = int(artist_id_raw) if artist_id_raw else None
     media = request.args.get("media")
     media_type = media if media in ("image", "video") else None
+    # newest/oldest key off effective_date (primary post / download); posted_new/
+    # posted_old off earliest_post_date (original publish across all posts). The
+    # default is posted_new so the grid leads with original publish date, not the
+    # download/repost the primary post points at (operator-flagged 2026-07-01).
     sort = request.args.get("sort")
-    sort = sort if sort in ("newest", "oldest") else "newest"
+    sort = sort if sort in ("newest", "oldest", "posted_new", "posted_old") else "posted_new"
     platform = request.args.get("platform") or None
     untagged = request.args.get("untagged") in ("1", "true", "yes")
     no_artist = request.args.get("no_artist") in ("1", "true", "yes")

backend/app/api/gpu.py

@@ -9,20 +9,25 @@ homelab admin.
 """
 
 import secrets
+from pathlib import Path
 
 from quart import Blueprint, jsonify, request
-from sqlalchemy import func, select
+from sqlalchemy import func, or_, select, update
 from sqlalchemy.dialects.postgresql import insert as pg_insert
 
 from ..extensions import get_session
 from ..models import AppSetting, GpuJob, ImageRecord, MLSettings
 from ..services.gallery_service import image_url
-from ..services.ml.embedder import MODEL_NAME as EMBED_MODEL_NAME
+from ..services.ml.gpu_jobs import GpuJobService, error_dedupe_statements
-from ..services.ml.gpu_jobs import GpuJobService
+from ..services.ml.gpu_triage import classify_reason, recover_defective_image
 from ..services.ml.regions import RegionService
 
 gpu_bp = Blueprint("gpu", __name__, url_prefix="/api/gpu")
 
+# Same container mount the maintenance tasks use (tasks/admin.py) — recovery
+# deletes the defective original + thumbnail under it.
+_IMAGES_ROOT = Path("/images")
+
 _TOKEN_KEY = "gpu_agent_token"
 
 
@@ -91,12 +96,152 @@ async def backfill():
     """Enqueue a job for every image that doesn't already have one for `task`."""
     body = await request.get_json(silent=True) or {}
     task = str(body.get("task") or "ccip")
-    from ..tasks.ml import enqueue_gpu_backfill
+    from ..tasks.gpu_queue import enqueue_gpu_backfill
 
     r = enqueue_gpu_backfill.delay(task)
     return jsonify({"celery_task_id": r.id, "task": task}), 202
 
 
+@gpu_bp.route("/reprocess", methods=["POST"])
+async def reprocess():
+    """Reset every done/error job of `task` back to pending so the agent re-runs
+    the WHOLE library under the current pipeline (e.g. after adding crop
+    detectors). Heavy — the back-catalogue is otherwise skipped by the backfills."""
+    body = await request.get_json(silent=True) or {}
+    task = str(body.get("task") or "ccip")
+    from ..tasks.gpu_queue import reprocess_gpu_jobs
+
+    r = reprocess_gpu_jobs.delay(task)
+    return jsonify({"celery_task_id": r.id, "task": task}), 202
+
+
+@gpu_bp.route("/retry_errors", methods=["POST"])
+async def retry_errors():
+    """Requeue every ERRORED job (all task types) back to pending — the scoped
+    recovery after an agent-side fix (e.g. the short-video sampler), where
+    /reprocess would needlessly re-run the whole done library too. Attempts and
+    the stored error reset so each job gets its full retry budget under the
+    fixed pipeline. Stale tombstones are pruned FIRST (loop-era duplicates and
+    rows a later success made moot — the same statements the backfills run), so
+    one failing file requeues as ONE job, never a fan-out of duplicates. Small
+    row count (errors only) → inline statements; the response carries the
+    counts for the UI toast. Triage-confirmed defects are NOT requeued (see
+    the WHERE below) — they stay on the recovery surface."""
+    async with get_session() as session:
+        pruned = 0
+        for stmt in error_dedupe_statements():
+            pruned += (await session.execute(stmt)).rowcount or 0
+        res = await session.execute(
+            update(GpuJob)
+            .where(
+                GpuJob.status == "error",
+                # Triage-confirmed DEFECTS stay errored: the integrity probe
+                # already proved the FILE is bad, so re-running the job just
+                # burns agent time re-minting the same tombstone — those go
+                # through /errors/<id>/recover instead.
+                or_(GpuJob.triage_status.is_(None),
+                    GpuJob.triage_status != "defect"),
+            )
+            .values(
+                status="pending", attempts=0, error=None, lease_token=None,
+                leased_at=None, lease_expires_at=None, triage_status=None,
+                updated_at=func.now(),
+            )
+        )
+        kept = (
+            await session.execute(
+                select(func.count()).select_from(GpuJob)
+                .where(GpuJob.status == "error")
+            )
+        ).scalar_one()
+        await session.commit()
+    return jsonify({
+        "requeued": res.rowcount or 0, "pruned": pruned, "defects_kept": kept,
+    })
+
+
+# --- Failure triage + recovery (#125) ------------------------------------
+
+@gpu_bp.route("/errors", methods=["GET"])
+async def errors():
+    """The triage view of the error tombstones: every errored job joined with
+    its image's integrity verdict, bucketed by reason for the overview. The
+    probe sweep (triage_gpu_errors, 15-min beat) fills triage_status; 'defect'
+    rows are the recovery surface's list."""
+    async with get_session() as session:
+        rows = (
+            await session.execute(
+                select(
+                    GpuJob.id, GpuJob.image_record_id, GpuJob.task,
+                    GpuJob.error, GpuJob.triage_status, GpuJob.updated_at,
+                    ImageRecord.integrity_status, ImageRecord.mime,
+                    ImageRecord.path, ImageRecord.thumbnail_path,
+                )
+                .join(ImageRecord, ImageRecord.id == GpuJob.image_record_id)
+                .where(GpuJob.status == "error")
+                .order_by(GpuJob.updated_at.desc())
+                .limit(500)
+            )
+        ).all()
+        total = (
+            await session.execute(
+                select(func.count()).select_from(GpuJob)
+                .where(GpuJob.status == "error")
+            )
+        ).scalar_one()
+    by_class: dict[str, int] = {}
+    triage = {"defect": 0, "file_ok": 0, "unclassified": 0}
+    items = []
+    for r in rows:
+        cls = classify_reason(r.error)
+        by_class[cls] = by_class.get(cls, 0) + 1
+        bucket = r.triage_status or "unclassified"
+        triage[bucket] = triage.get(bucket, 0) + 1
+        items.append({
+            "job_id": r.id,
+            "image_id": r.image_record_id,
+            "task": r.task,
+            "error": r.error,
+            "reason_class": cls,
+            "triage_status": r.triage_status,
+            "integrity_status": r.integrity_status,
+            "mime": r.mime,
+            "image_url": image_url(r.path),
+            "thumbnail_url": (
+                image_url(r.thumbnail_path) if r.thumbnail_path else None
+            ),
+            "updated_at": r.updated_at.isoformat() if r.updated_at else None,
+        })
+    return jsonify({
+        "total": total, "by_class": by_class, "triage": triage, "items": items,
+    })
+
+
+@gpu_bp.route("/errors/triage", methods=["POST"])
+async def errors_triage():
+    """Run the probe sweep NOW (the card's button) instead of waiting out the
+    15-minute beat cadence."""
+    from ..tasks.maintenance import triage_gpu_errors
+
+    r = triage_gpu_errors.delay()
+    return jsonify({"celery_task_id": r.id}), 202
+
+
+@gpu_bp.route("/errors/<int:image_id>/recover", methods=["POST"])
+async def errors_recover(image_id: int):
+    """Recover a defect-triaged original: delete the bad copy + record and
+    re-poll its subscription Source (a fresh fetch re-imports the file, which
+    re-enters the GPU pipeline). Returns status 'no_source' when nothing
+    pollable resolves — the file needs manual replacement there."""
+    async with get_session() as session:
+        result = await session.run_sync(
+            lambda s: recover_defective_image(
+                s, image_id, images_root=_IMAGES_ROOT,
+            )
+        )
+    return jsonify(result)
+
+
 # --- Agent (bearer token): lease / submit / heartbeat / fail ------------
 
 @gpu_bp.route("/jobs/lease", methods=["POST"])
@@ -138,11 +283,12 @@ async def lease():
                 # For video/animated: the agent samples at this cadence.
                 "frame_interval_seconds": ml.video_frame_interval_seconds,
                 "max_frames": ml.video_max_frames,
-                # The embedding model the agent must use for concept crops, so
+                # The embedding model the agent must use for concept crops + the
-                # its region vectors land in the SAME space the heads trained in.
+                # whole-image 'embed' task, so its vectors land in the SAME space
-                # Server-announced → the agent stays model-agnostic; a swap is a
+                # the heads trained in. Server-announced FROM THE SETTING → the
-                # server setting + a re-embed migration, never an agent change.
+                # agent stays model-agnostic; an operator swap is a setting + a
-                "embed_model_name": EMBED_MODEL_NAME,
+                # re-embed, never an agent change.
+                "embed_model_name": ml.embedder_model_name,
                 "embed_version": ml.embedder_model_version,
             })
     return jsonify({"jobs": out})
@@ -188,6 +334,34 @@ async def submit():
     return jsonify({"ok": True, "stored": len(regions)})
 
 
+@gpu_bp.route("/jobs/submit_embedding", methods=["POST"])
+async def submit_embedding():
+    """Store a whole-image SigLIP embedding (the 'embed' task) on image_record +
+    close the job. Body: {agent_id, job_id, embedding:[...], embedding_version}.
+    This is how the GPU agent re-embeds the library under a new model (#1190) —
+    much faster than the CPU ml-worker at higher resolutions."""
+    body = await request.get_json(silent=True) or {}
+    agent_id = str(body.get("agent_id") or "agent")
+    job_id = body.get("job_id")
+    embedding = body.get("embedding")
+    version = body.get("embedding_version")
+    if job_id is None or not embedding or not version:
+        return jsonify({"error": "job_id, embedding, embedding_version required"}), 400
+    async with get_session() as session:
+        if not await _agent_authed(session):
+            return jsonify({"error": "unauthorized"}), 401
+        job = await session.get(GpuJob, int(job_id))
+        if job is None or job.status != "leased" or job.lease_token != agent_id:
+            return jsonify({"error": "lease_invalid"}), 409
+        img = await session.get(ImageRecord, job.image_record_id)
+        if img is not None:
+            img.siglip_embedding = embedding
+            img.siglip_model_version = version
+        await GpuJobService(session).complete(agent_id, int(job_id))
+        await session.commit()
+    return jsonify({"ok": True})
+
+
 @gpu_bp.route("/jobs/fail", methods=["POST"])
 async def fail():
     body = await request.get_json(silent=True) or {}

backend/app/api/ml_admin.py

@@ -1,4 +1,4 @@
-"""ML admin API: settings, backfill trigger, centroid recompute trigger."""
+"""ML admin API: settings + backfill trigger."""
 
 from quart import Blueprint, jsonify, request
 
@@ -9,14 +9,9 @@ ml_admin_bp = Blueprint("ml_admin", __name__, url_prefix="/api/ml")
 
 
 _EDITABLE = (
-    "suggestion_threshold_character",
+    "cpu_embed_enabled",
-    "suggestion_threshold_general",
-    "centroid_similarity_threshold",
-    "min_reference_images",
-    "tagger_store_floor",
     "video_frame_interval_seconds",
     "video_max_frames",
-    "video_min_tag_frames",
     "head_min_positives",
     "head_auto_apply_precision",
     "head_auto_apply_enabled",
@@ -24,9 +19,41 @@ _EDITABLE = (
     "ccip_match_threshold",
     "ccip_auto_apply_enabled",
     "ccip_auto_apply_threshold",
+    "embedder_model_name",
+    "embedder_model_version",
 )
 
 
+# Supported embedders for the Settings dropdown — all 1152-d so a swap is a
+# drop-in (re-embed + retrain, no schema change). Server-authoritative so the UI
+# never free-types a model name.
+SUPPORTED_EMBEDDERS = (
+    {
+        "name": "google/siglip2-so400m-patch16-512",
+        "version": "siglip2-so400m-patch16-512",
+        "label": "SigLIP 2 · so400m · 512px (recommended)",
+        "dim": 1152,
+    },
+    {
+        "name": "google/siglip2-so400m-patch16-384",
+        "version": "siglip2-so400m-patch16-384",
+        "label": "SigLIP 2 · so400m · 384px (faster)",
+        "dim": 1152,
+    },
+    {
+        "name": "google/siglip-so400m-patch14-384",
+        "version": "siglip-so400m-patch14-384",
+        "label": "SigLIP 1 · so400m · 384px (original)",
+        "dim": 1152,
+    },
+)
+
+
+@ml_admin_bp.route("/embedder-models", methods=["GET"])
+async def embedder_models():
+    return jsonify({"models": list(SUPPORTED_EMBEDDERS)})
+
+
 @ml_admin_bp.route("/settings", methods=["GET"])
 async def get_settings():
     from sqlalchemy import select
@@ -37,15 +64,9 @@ async def get_settings():
         ).scalar_one()
     return jsonify(
         {
-            "suggestion_threshold_character": s.suggestion_threshold_character,
+            "cpu_embed_enabled": s.cpu_embed_enabled,
-            "suggestion_threshold_general": s.suggestion_threshold_general,
-            "centroid_similarity_threshold": s.centroid_similarity_threshold,
-            "min_reference_images": s.min_reference_images,
-            "tagger_store_floor": s.tagger_store_floor,
             "video_frame_interval_seconds": s.video_frame_interval_seconds,
             "video_max_frames": s.video_max_frames,
-            "video_min_tag_frames": s.video_min_tag_frames,
-            "tagger_model_version": s.tagger_model_version,
             "embedder_model_version": s.embedder_model_version,
             "head_min_positives": s.head_min_positives,
             "head_auto_apply_precision": s.head_auto_apply_precision,
@@ -54,6 +75,7 @@ async def get_settings():
             "ccip_match_threshold": s.ccip_match_threshold,
             "ccip_auto_apply_enabled": s.ccip_auto_apply_enabled,
             "ccip_auto_apply_threshold": s.ccip_auto_apply_threshold,
+            "embedder_model_name": s.embedder_model_name,
         }
     )
 
@@ -89,31 +111,12 @@ async def patch_settings():
 
 
 def _validate(p: dict) -> str | None:
-    """Returns an error string if the proposed settings are invalid, else None.
+    """Returns an error string if the proposed settings are invalid, else None."""
-
+    # Video embedding (#747).
-    Invariant (plan-task #764): the per-category suggestion thresholds can't
-    drop below tagger_store_floor — nothing below the floor is stored, so a
-    lower threshold would silently surface nothing in that gap. The UI clamps
-    the sliders to the floor; this is the server-side backstop.
-    """
-    floor = p["tagger_store_floor"]
-    if not (0.0 <= floor <= 1.0):
-        return "tagger_store_floor must be between 0 and 1"
-    for cat in ("character", "general"):
-        if p[f"suggestion_threshold_{cat}"] < floor:
-            return (
-                f"suggestion_threshold_{cat} cannot be below tagger_store_floor "
-                f"({floor}) — predictions below the floor are not stored"
-            )
-    # Video tagging (#747).
     if p["video_frame_interval_seconds"] <= 0:
         return "video_frame_interval_seconds must be > 0"
     if p["video_max_frames"] < 1:
         return "video_max_frames must be >= 1"
-    if p["video_min_tag_frames"] < 1:
-        return "video_min_tag_frames must be >= 1"
-    if p["video_min_tag_frames"] > p["video_max_frames"]:
-        return "video_min_tag_frames cannot exceed video_max_frames"
     # Head training (#114).
     if int(p["head_min_positives"]) < 1:
         return "head_min_positives must be >= 1"
@@ -125,6 +128,11 @@ def _validate(p: dict) -> str | None:
         return "ccip_match_threshold must be between 0.5 and 0.999"
     if not (0.5 <= float(p["ccip_auto_apply_threshold"]) <= 0.999):
         return "ccip_auto_apply_threshold must be between 0.5 and 0.999"
+    # Embedder model swap (#1190): both must be non-empty. Changing them means a
+    # different embedding space — the operator must re-embed + retrain after.
+    for key in ("embedder_model_name", "embedder_model_version"):
+        if not str(p[key]).strip():
+            return f"{key} must not be empty"
     return None
 
 
@@ -134,11 +142,3 @@ async def trigger_backfill():
 
     r = backfill.delay()
     return jsonify({"celery_task_id": r.id}), 202
-
-
-@ml_admin_bp.route("/recompute-centroids", methods=["POST"])
-async def trigger_recompute():
-    from ..tasks.ml import recompute_centroids
-
-    r = recompute_centroids.delay()
-    return jsonify({"celery_task_id": r.id}), 202

backend/app/api/suggestions.py

@@ -3,31 +3,12 @@
 from quart import Blueprint, jsonify, request
 
 from ..extensions import get_session
-from ..models import Tag, TagAllowlist
 from ..services.ml.allowlist import AllowlistService
 from ..services.ml.suggestions import SuggestionService
 
 suggestions_bp = Blueprint("suggestions", __name__, url_prefix="/api")
 
 
-async def _accept_payload(session, svc, newly_added: bool, tag_id: int) -> dict:
-    """Shape the accept/alias response. When accepting newly allowlists a tag,
-    include the coverage PROJECTION (at the tag's threshold) so the UI can show
-    a non-blocking "auto-applying to ~N images" toast — the actual apply runs
-    async via apply_allowlist_tags, so this is an estimate, not a post-hoc
-    count (#7)."""
-    payload = {"allowlisted": newly_added}
-    if newly_added:
-        tag = await session.get(Tag, tag_id)
-        row = await session.get(TagAllowlist, tag_id)
-        payload["tag_id"] = tag_id
-        payload["tag_name"] = tag.name if tag is not None else None
-        payload["projected_count"] = await svc.coverage(
-            tag_id, row.min_confidence if row is not None else 0.90,
-        )
-    return payload
-
-
 @suggestions_bp.route("/images/<int:image_id>/suggestions", methods=["GET"])
 async def get_suggestions(image_id: int):
     # ?min=<float> overrides the configured per-category thresholds so the typed
@@ -83,15 +64,9 @@ async def accept_suggestion(image_id: int):
         return jsonify({"error": "tag_id required"}), 400
     tag_id = body["tag_id"]
     async with get_session() as session:
-        svc = AllowlistService(session)
+        await AllowlistService(session).accept(image_id, tag_id)
-        newly_added = await svc.accept(image_id, tag_id)
-        payload = await _accept_payload(session, svc, newly_added, tag_id)
         await session.commit()
-    if newly_added:
+    return jsonify({"accepted": True, "tag_id": tag_id})
-        from ..tasks.ml import apply_allowlist_tags
-
-        apply_allowlist_tags.delay(tag_id=tag_id)
-    return jsonify(payload)
 
 
 @suggestions_bp.route(
@@ -104,22 +79,14 @@ async def alias_suggestion(image_id: int):
         return jsonify({"error": f"required: {sorted(required)}"}), 400
     canonical_tag_id = body["canonical_tag_id"]
     async with get_session() as session:
-        svc = AllowlistService(session)
+        await AllowlistService(session).add_alias_and_accept(
-        newly_added = await svc.add_alias_and_accept(
             image_id,
             body["alias_string"],
             body["alias_category"],
             canonical_tag_id,
         )
-        payload = await _accept_payload(
-            session, svc, newly_added, canonical_tag_id,
-        )
         await session.commit()
-    if newly_added:
+    return jsonify({"accepted": True, "tag_id": canonical_tag_id})
-        from ..tasks.ml import apply_allowlist_tags
-
-        apply_allowlist_tags.delay(tag_id=canonical_tag_id)
-    return jsonify(payload)
 
 
 @suggestions_bp.route(

backend/app/api/tag_eval.py

@@ -1,70 +0,0 @@
-"""Tag-eval API (#1130): trigger + revisit the head-vs-centroid eval.
-
-The run + full report live in the tag_eval_run row, so the admin card rehydrates
-from GET (history / detail) on mount — the report survives navigation rather than
-living in transient frontend state.
-"""
-
-from quart import Blueprint, jsonify, request
-from sqlalchemy import select
-
-from ..extensions import get_session
-from ..models import TagEvalRun
-from ..services.ml.tag_eval import EvalAlreadyRunning, start_tag_eval_run
-
-tag_eval_bp = Blueprint("tag_eval", __name__, url_prefix="/api/tag-eval")
-
-
-def _serialize(run: TagEvalRun, *, include_report: bool) -> dict:
-    out = {
-        "id": run.id,
-        "params": run.params,
-        "status": run.status,
-        "started_at": run.started_at.isoformat() if run.started_at else None,
-        "finished_at": run.finished_at.isoformat() if run.finished_at else None,
-        "error": run.error,
-    }
-    if include_report:
-        out["report"] = run.report
-    return out
-
-
-@tag_eval_bp.route("", methods=["POST"])
-async def create():
-    body = await request.get_json(silent=True) or {}
-    params = body.get("params") or body or {}
-    async with get_session() as session:
-        try:
-            run_id = await session.run_sync(
-                lambda s: start_tag_eval_run(s, params)
-            )
-        except EvalAlreadyRunning as running:
-            return jsonify({
-                "error": "eval_already_running",
-                "running_id": int(running.args[0]),
-            }), 409
-        await session.commit()
-    return jsonify({"run_id": run_id, "status": "running"}), 202
-
-
-@tag_eval_bp.route("", methods=["GET"])
-async def history():
-    try:
-        limit = min(int(request.args.get("limit", "20")), 100)
-    except ValueError:
-        return jsonify({"error": "invalid_limit"}), 400
-    async with get_session() as session:
-        rows = (await session.execute(
-            select(TagEvalRun).order_by(TagEvalRun.id.desc()).limit(limit)
-        )).scalars().all()
-        # List is light — no full report (the detail endpoint carries it).
-        return jsonify({"runs": [_serialize(r, include_report=False) for r in rows]})
-
-
-@tag_eval_bp.route("/<int:run_id>", methods=["GET"])
-async def detail(run_id: int):
-    async with get_session() as session:
-        run = await session.get(TagEvalRun, run_id)
-        if run is None:
-            return jsonify({"error": "not_found"}), 404
-        return jsonify(_serialize(run, include_report=True))

backend/app/api/tags.py

@@ -1,13 +1,14 @@
 """Tags API: autocomplete, create, list/add/remove for an image."""
 
 from quart import Blueprint, jsonify, request
-from sqlalchemy import exists, select
+from sqlalchemy import func, select
 from sqlalchemy.dialects.postgresql import insert as pg_insert
 from sqlalchemy.exc import IntegrityError
 
 from ..extensions import get_session
-from ..models import Tag, TagKind, TagPositiveConfirmation
+from ..models import Tag, TagHead, TagKind, TagPositiveConfirmation
-from ..models.tag_allowlist import TagAllowlist
+from ..models.tag import image_tag
+from ..models.tag_suggestion_rejection import TagSuggestionRejection
 from ..services.bulk_tag_service import BulkTagService
 from ..services.ml.aliases import AliasService
 from ..services.series_match_service import SeriesMatchService
@@ -61,6 +62,117 @@ def _parse_bulk_ids(
     return ids, None
 
 
+# Application-source groupings (image_tag.source). HUMAN = operator signal;
+# AUTO = machine-applied (heads/CCIP, + legacy Camie ml_auto).
+_SOURCE_GROUPS = {
+    "human": ("manual", "ml_accepted"),
+    "manual": ("manual",),
+    "accepted": ("ml_accepted",),
+    "auto": ("head_auto", "ccip_auto", "ml_auto"),
+}
+
+
+@tags_bp.route("/tags/top", methods=["GET"])
+async def tags_top():
+    """Top tags by image count — a fast indexed aggregate for ANALYSIS (not the
+    paged UI directory, which is alphabetical + builds previews). Params:
+    ?kind=general|character|fandom|… ?source=all|human|manual|accepted|auto
+    ?limit=50 (cap 500) ?min_count=N. → {tags:[{tag_id,name,kind,count}]} desc."""
+    kind = _coerce_kind(request.args.get("kind"))
+    try:
+        limit = min(max(int(request.args.get("limit", "50")), 1), 500)
+    except ValueError:
+        return jsonify({"error": "limit must be an integer"}), 400
+    min_count = None
+    if "min_count" in request.args:
+        try:
+            min_count = int(request.args["min_count"])
+        except ValueError:
+            return jsonify({"error": "min_count must be an integer"}), 400
+    src_vals = _SOURCE_GROUPS.get((request.args.get("source") or "all").lower())
+
+    cnt = func.count(image_tag.c.image_record_id)
+    stmt = (
+        select(Tag.id, Tag.name, Tag.kind, cnt.label("count"))
+        .select_from(Tag)
+        .join(image_tag, image_tag.c.tag_id == Tag.id)
+        .group_by(Tag.id, Tag.name, Tag.kind)
+        .order_by(cnt.desc(), Tag.name.asc())
+        .limit(limit)
+    )
+    if kind is not None:
+        stmt = stmt.where(Tag.kind == kind)
+    if src_vals is not None:
+        stmt = stmt.where(image_tag.c.source.in_(src_vals))
+    if min_count is not None:
+        stmt = stmt.having(cnt >= min_count)
+    async with get_session() as session:
+        rows = (await session.execute(stmt)).all()
+    return jsonify({"tags": [
+        {
+            "tag_id": r.id, "name": r.name,
+            "kind": r.kind.value if hasattr(r.kind, "value") else str(r.kind),
+            "count": r.count,
+        }
+        for r in rows
+    ]})
+
+
+@tags_bp.route("/tags/<int:tag_id>/stats", methods=["GET"])
+async def tag_stats(tag_id: int):
+    """Per-tag dataset health: total + per-source application counts (human vs
+    machine), rejection count, and whether a trained head exists. Read-only,
+    analysis-shaped — backs concept-readiness + source-split decisions."""
+    async with get_session() as session:
+        tag = await session.get(Tag, tag_id)
+        if tag is None:
+            return jsonify({"error": "not found"}), 404
+        by_source = dict(
+            (
+                await session.execute(
+                    select(image_tag.c.source, func.count())
+                    .where(image_tag.c.tag_id == tag_id)
+                    .group_by(image_tag.c.source)
+                )
+            ).all()
+        )
+        rejected = (
+            await session.execute(
+                select(func.count())
+                .select_from(TagSuggestionRejection)
+                .where(TagSuggestionRejection.tag_id == tag_id)
+            )
+        ).scalar_one()
+        has_head = (
+            await session.execute(
+                select(func.count())
+                .select_from(TagHead)
+                .where(TagHead.tag_id == tag_id)
+            )
+        ).scalar_one() > 0
+    human = by_source.get("manual", 0) + by_source.get("ml_accepted", 0)
+    auto = (
+        by_source.get("head_auto", 0)
+        + by_source.get("ccip_auto", 0)
+        + by_source.get("ml_auto", 0)
+    )
+    return jsonify({
+        "tag_id": tag_id,
+        "name": tag.name,
+        "kind": tag.kind.value if hasattr(tag.kind, "value") else str(tag.kind),
+        "count_total": sum(by_source.values()),
+        "count_human": human,
+        "count_manual": by_source.get("manual", 0),
+        "count_accepted": by_source.get("ml_accepted", 0),
+        "count_auto": auto,
+        "count_head_auto": by_source.get("head_auto", 0),
+        "count_ccip_auto": by_source.get("ccip_auto", 0),
+        "count_rejected": rejected,
+        "by_source": by_source,
+        "has_head": has_head,
+    })
+
+
 @tags_bp.route("/tags/autocomplete", methods=["GET"])
 async def autocomplete():
     q = request.args.get("q", "")
@@ -297,19 +409,6 @@ async def merge_tag(source_id: int):
             status = 404 if "not found" in msg else 400
             return jsonify({"error": msg}), status
         await session.commit()
-        target_allowlisted = await session.scalar(
-            select(exists().where(TagAllowlist.tag_id == result.target_id))
-        )
-    if target_allowlisted:
-        from ..tasks.ml import apply_allowlist_tags
-
-        apply_allowlist_tags.delay(tag_id=result.target_id)
-    # Tag merge invalidates the target's centroid (the merged-in source
-    # tag's images now contribute to it). Daily list_drifted catches it
-    # within 24h, but eager recompute closes the suggestion-quality dip
-    # in the meantime. Audit 2026-06-02.
-    from ..tasks.ml import recompute_centroid
-    recompute_centroid.delay(result.target_id)
     return jsonify(
         {
             "target": {

backend/app/celery_app.py

@@ -7,7 +7,7 @@ Queues:
   download    — gallery-dl tasks (FC-3)
   scan        — periodic source checks (FC-3) — kept separate so long imports
                 don't starve the scheduler
-  maintenance — pHash recomputation, centroid rebuild, etc. (FC-2/FC-3)
+  maintenance — recovery sweeps, pHash backfill, GPU-queue coordination, etc.
   default     — anything not explicitly routed
 """
 
@@ -29,6 +29,7 @@ def make_celery() -> Celery:
             "backend.app.tasks.thumbnail",
             "backend.app.tasks.maintenance",
             "backend.app.tasks.ml",
+            "backend.app.tasks.gpu_queue",
             "backend.app.tasks.download",
             "backend.app.tasks.external",
             "backend.app.tasks.backup",
@@ -41,6 +42,11 @@ def make_celery() -> Celery:
         task_routes={
             "backend.app.tasks.import_file.*": {"queue": "import"},
             "backend.app.tasks.ml.*": {"queue": "ml"},
+            # GPU-queue coordination (backfill enqueues, orphan recovery,
+            # reprocess) is pure DB work — it rides the maintenance quick lane
+            # so the GPU agent pipeline works even on stacks that drop the
+            # (now-optional, B3) ml-worker container entirely.
+            "backend.app.tasks.gpu_queue.*": {"queue": "maintenance"},
             "backend.app.tasks.thumbnail.*": {"queue": "thumbnail"},
             "backend.app.tasks.download.*": {"queue": "download"},
             # External file-host fetches are downloads — same lane (they can run
@@ -97,18 +103,6 @@ def make_celery() -> Celery:
                 "task": "backend.app.tasks.maintenance.cleanup_old_tasks",
                 "schedule": 86400.0,  # daily
             },
-            "ml-backfill-daily": {
-                "task": "backend.app.tasks.ml.backfill",
-                "schedule": 86400.0,
-            },
-            "recompute-centroids-daily": {
-                "task": "backend.app.tasks.ml.recompute_centroids",
-                "schedule": 86400.0,
-            },
-            "apply-allowlist-sweep-daily": {
-                "task": "backend.app.tasks.ml.apply_allowlist_tags",
-                "schedule": 86400.0,
-            },
             "train-heads-nightly": {
                 "task": "backend.app.tasks.ml.scheduled_train_heads",
                 "schedule": 86400.0,  # passive cadence; manual retrain stays available
@@ -118,18 +112,27 @@ def make_celery() -> Celery:
                 "schedule": 86400.0,  # no-op unless head_auto_apply_enabled
             },
             "recover-orphaned-gpu-jobs": {
-                "task": "backend.app.tasks.ml.recover_orphaned_gpu_jobs",
+                "task": "backend.app.tasks.gpu_queue.recover_orphaned_gpu_jobs",
                 "schedule": 60.0,  # quick pickup of work a dead agent orphaned
             },
+            "triage-gpu-errors": {
+                "task": "backend.app.tasks.maintenance.triage_gpu_errors",
+                "schedule": 900.0,  # probe errored jobs' files → defect/file_ok
+            },
             "enqueue-ccip-backfill-hourly": {
-                "task": "backend.app.tasks.ml.enqueue_gpu_backfill",
+                "task": "backend.app.tasks.gpu_queue.enqueue_gpu_backfill",
-                "schedule": 3600.0,  # auto-feed new images (+ retry errored) so
+                "schedule": 3600.0,  # auto-feed NEW images; errored are
-                "args": ("ccip",),   # the queue keeps moving without the button
+                "args": ("ccip",),   # tombstoned — retry is the button only
             },
             "enqueue-siglip-backfill-daily": {
-                "task": "backend.app.tasks.ml.enqueue_gpu_backfill",
+                "task": "backend.app.tasks.gpu_queue.enqueue_gpu_backfill",
-                "schedule": 86400.0,  # drain the concept-crop back-catalogue +
+                "schedule": 86400.0,  # drain the concept-crop back-catalogue
-                "args": ("siglip",),  # retry failed embeds, no button needed
+                "args": ("siglip",),  # (errored are tombstoned, not retried)
+            },
+            "enqueue-embed-backfill-daily": {
+                "task": "backend.app.tasks.gpu_queue.enqueue_gpu_backfill",
+                "schedule": 86400.0,  # whole-image re-embed under the current
+                "args": ("embed",),   # model (an operator swap) drains via agent
             },
             "ccip-auto-apply-daily": {
                 "task": "backend.app.tasks.ml.scheduled_ccip_auto_apply",
@@ -186,10 +189,6 @@ def make_celery() -> Celery:
                 "task": "backend.app.tasks.maintenance.recover_stalled_library_audit_runs",
                 "schedule": 300.0,
             },
-            "recover-stalled-tag-eval-runs": {
-                "task": "backend.app.tasks.maintenance.recover_stalled_tag_eval_runs",
-                "schedule": 300.0,
-            },
             "recover-stalled-head-training-runs": {
                 "task": "backend.app.tasks.maintenance.recover_stalled_head_training_runs",
                 "schedule": 300.0,

backend/app/models/__init__.py

@@ -13,7 +13,6 @@ from .head_auto_apply_run import HeadAutoApplyRun
 from .head_metric import HeadMetric
 from .head_metrics_snapshot import HeadMetricsSnapshot
 from .head_training_run import HeadTrainingRun
-from .image_prediction import ImagePrediction
 from .image_provenance import ImageProvenance
 from .image_record import ImageRecord
 from .image_region import ImageRegion
@@ -34,11 +33,8 @@ from .subscribestar_failed_media import SubscribeStarFailedMedia
 from .subscribestar_seen_media import SubscribeStarSeenMedia
 from .tag import Tag, TagKind, image_tag
 from .tag_alias import TagAlias
-from .tag_allowlist import TagAllowlist
-from .tag_eval_run import TagEvalRun
 from .tag_head import TagHead
 from .tag_positive_confirmation import TagPositiveConfirmation
-from .tag_reference_embedding import TagReferenceEmbedding
 from .tag_suggestion_rejection import TagSuggestionRejection
 from .task_run import TaskRun
 
@@ -60,7 +56,6 @@ __all__ = [
     "SeriesPage",
     "SeriesSuggestion",
     "ImageRecord",
-    "ImagePrediction",
     "ImageProvenance",
     "ImageRegion",
     "Tag",
@@ -79,11 +74,8 @@ __all__ = [
     "HeadMetricsSnapshot",
     "HeadTrainingRun",
     "TagAlias",
-    "TagAllowlist",
-    "TagEvalRun",
     "TagHead",
     "TagPositiveConfirmation",
-    "TagReferenceEmbedding",
     "TagSuggestionRejection",
     "TaskRun",
 ]

backend/app/models/gpu_job.py

@@ -14,7 +14,16 @@ pending for another agent).
 
 from datetime import datetime
 
-from sqlalchemy import DateTime, ForeignKey, Integer, String, Text, func
+from sqlalchemy import (
+    DateTime,
+    ForeignKey,
+    Index,
+    Integer,
+    String,
+    Text,
+    func,
+    text,
+)
 from sqlalchemy.orm import Mapped, mapped_column
 
 from .base import Base
@@ -23,6 +32,17 @@ from .base import Base
 class GpuJob(Base):
     __tablename__ = "gpu_job"
 
+    # Partial indexes over just the live slice (see migration 0070): the lease
+    # reads the lowest-id pending jobs on the hot path, and reclaims expired
+    # leases as a backstop — both stay O(batch) as done/error history grows.
+    __table_args__ = (
+        Index("ix_gpu_job_pending", "id", postgresql_where=text("status = 'pending'")),
+        Index(
+            "ix_gpu_job_leased_expires", "lease_expires_at",
+            postgresql_where=text("status = 'leased'"),
+        ),
+    )
+
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
     image_record_id: Mapped[int] = mapped_column(
         ForeignKey("image_record.id", ondelete="CASCADE"), index=True
@@ -42,6 +62,11 @@ class GpuJob(Base):
     )
     attempts: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
     error: Mapped[str | None] = mapped_column(Text, nullable=True)
+    # Triage verdict for an ERRORED job (#125): NULL = not yet probed;
+    # 'defect' = the integrity probe says the FILE itself is bad (surfaced for
+    # recovery, excluded from /retry_errors); 'file_ok' = the file passes —
+    # the failure was operational (timeout/transient), safe to retry.
+    triage_status: Mapped[str | None] = mapped_column(String(16), nullable=True)
     created_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()
     )

backend/app/models/head_training_run.py

@@ -1,7 +1,7 @@
 """HeadTrainingRun — persisted lifecycle of a head-training batch (#114).
 
-Mirrors TagEvalRun so the run SURVIVES navigation and the admin card can show
+A persisted run row (not transient frontend state) so the run SURVIVES
-live + historical status instead of holding it in transient frontend state.
+navigation and the admin card can show live + historical status.
 Training is idempotent (it upserts tag_head rows), so a SIGKILL'd run is harmless
 — a maintenance recovery sweep flips a stalled `running` row to `error`, and the
 next run re-trains. State machine: running → ready / error.
@@ -37,8 +37,8 @@ class HeadTrainingRun(Base):
     n_trained: Mapped[int | None] = mapped_column(Integer, nullable=True)
     n_skipped: Mapped[int | None] = mapped_column(Integer, nullable=True)
     error: Mapped[str | None] = mapped_column(Text, nullable=True)
-    # Last time the task made progress — the recovery sweep tells a live run from
+    # Last time the task made progress — the recovery sweep tells a live run
-    # a SIGKILL'd one by this (mirrors TagEvalRun).
+    # from a SIGKILL'd one by this.
     last_progress_at: Mapped[datetime | None] = mapped_column(
         DateTime(timezone=True), nullable=True
     )

backend/app/models/image_prediction.py

@@ -1,37 +0,0 @@
-"""ImagePrediction — one row per (image, tagger vocab prediction).
-
-Replaces the image_record.tagger_predictions JSON blob (#768). Storing the
-raw Camie/booru vocab name (not a tag_id) preserves the suggestion read
-path's semantics: raw_name → canonical Tag resolution happens at read time
-via the alias map, and accepting a prediction can CREATE the Tag. The store
-floor (ml_settings.tagger_store_floor) is applied at WRITE time, so only
-predictions >= the floor land here.
-"""
-
-from sqlalchemy import Float, ForeignKey, Index, String, UniqueConstraint
-from sqlalchemy.orm import Mapped, mapped_column
-
-from .base import Base
-
-
-class ImagePrediction(Base):
-    __tablename__ = "image_prediction"
-    __table_args__ = (
-        UniqueConstraint(
-            "image_record_id", "raw_name", name="image_raw_name",
-        ),
-        # Per-image read (suggestion build) and the "images with tag X above
-        # Y" query the JSON blob never allowed.
-        Index("ix_image_prediction_image", "image_record_id"),
-        Index("ix_image_prediction_name_score", "raw_name", "score"),
-    )
-
-    id: Mapped[int] = mapped_column(primary_key=True)
-    image_record_id: Mapped[int] = mapped_column(
-        ForeignKey("image_record.id", ondelete="CASCADE"), nullable=False,
-    )
-    # The raw tagger vocab key (booru form) — NOT a tag_id. Resolved to a
-    # canonical Tag at read time, exactly as the old JSON keys were.
-    raw_name: Mapped[str] = mapped_column(String(255), nullable=False)
-    category: Mapped[str] = mapped_column(String(64), nullable=False)
-    score: Mapped[float] = mapped_column(Float, nullable=False)

backend/app/models/image_record.py

@@ -9,7 +9,6 @@ from datetime import datetime
 
 from pgvector.sqlalchemy import Vector
 from sqlalchemy import (
-    JSON,
     BigInteger,
     DateTime,
     Enum,
@@ -77,19 +76,13 @@ class ImageRecord(Base):
         ForeignKey("artist.id", ondelete="SET NULL"), nullable=True, index=True
     )
 
-    # ML fields (populated by FC-2's ml-worker). Per-tag predictions live in the
+    # ML fields (populated by the ml-worker / GPU agent). 1152 = SigLIP-so400m
-    # normalized image_prediction table (#768) — the tagger_predictions JSON
+    # embedding dim; siglip_model_version stamps which model produced it (so an
-    # column was dropped in migration 0046. tagger_model_version stays as the
+    # operator model swap, #1190, can re-embed the stale rows). A different-dim
-    # "has this been tagged / is it current?" signal the backfill sweep reads.
+    # model would need a column-width migration.
-    tagger_model_version: Mapped[str | None] = mapped_column(String(128), nullable=True)
-    # 1152 = SigLIP-so400m embedding dim. Swapping models in FC-2 may require
-    # a column-width migration.
     siglip_embedding: Mapped[list[float] | None] = mapped_column(Vector(1152), nullable=True)
     siglip_model_version: Mapped[str | None] = mapped_column(String(128), nullable=True)
 
-    # Centroid score cache (populated post-tagging)
-    centroid_scores: Mapped[dict | None] = mapped_column(JSON, nullable=True)
-
     created_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()
     )
@@ -104,6 +97,16 @@ class ImageRecord(Base):
     effective_date: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()
     )
+    # Denormalized ORIGINAL-publish sort key (alembic 0071) = MIN(post_date)
+    # across ALL of the image's provenance posts, else created_at. effective_date
+    # above keys off the PRIMARY post (often the repost/download the file came
+    # from); this keys off the earliest publish across EVERY post the image
+    # appears in, so the gallery can sort by when content was first posted rather
+    # than when it was downloaded (operator-flagged 2026-07-01). Maintained by
+    # services/importer.py, recomputed whenever a dated post is linked.
+    earliest_post_date: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )
     updated_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True),
         nullable=False,

backend/app/models/image_region.py

@@ -31,7 +31,10 @@ class ImageRegion(Base):
         ForeignKey("image_record.id", ondelete="CASCADE"), index=True
     )
     # 'frame' (a whole video frame → SigLIP bag) | 'face' | 'figure' (→ CCIP
-    # character id) | 'concept' (→ SigLIP head bag).
+    # character id) | 'concept' (→ SigLIP head bag) | 'panel' (a comic panel crop,
+    # also SigLIP → the bag). Free String, not an enum — proposers can add kinds
+    # without a migration; the bag scorer keys on a non-null siglip_embedding, not
+    # the kind, so any SigLIP-embedded region joins the bag.
     kind: Mapped[str] = mapped_column(String(16), nullable=False)
     # For video/animated media: the source frame's timestamp in SECONDS. NULL for
     # static images. Lets a video be a BAG of per-frame instances (fixes the

backend/app/models/ml_settings.py

@@ -23,46 +23,25 @@ class MLSettings(Base):
     __table_args__ = (CheckConstraint("id = 1", name="singleton"),)
 
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
-    suggestion_threshold_character: Mapped[float] = mapped_column(
+    # CPU whole-image embedding (B3, operator 2026-07-02). The ml-worker's ONLY
-        Float, nullable=False, default=0.70
+    # processing role is the embed fallback for stacks WITHOUT a GPU agent — ON
+    # by default so a fresh install works with no agent. Stacks that run the
+    # agent and drop the ml-worker container turn this OFF so import hooks stop
+    # queueing embed work nothing will consume (the daily GPU 'embed' backfill
+    # covers those images instead).
+    cpu_embed_enabled: Mapped[bool] = mapped_column(
+        Boolean, nullable=False, default=True
     )
-    # Default raised 0.50 → 0.70 on 2026-06-02 — operator-flagged 0.50
+    # Video embedding (#747). Sample one frame every N seconds (fixed CADENCE, not
-    # surfaced too many low-confidence picks; 0.70 keeps the rail
+    # a fixed count) so coverage reflects real screen time regardless of length;
-    # signal-rich while still surfacing more than the original 0.95
+    # cap the total so a long video can't explode into hundreds of embeds. The
-    # which hid almost everything. Operator-tunable via Settings → ML.
+    # per-frame SigLIP embeddings are mean-pooled. Operator-tunable.
-    suggestion_threshold_general: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.70
-    )
-    centroid_similarity_threshold: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.55
-    )
-    # Ingest floor: tagger predictions below this confidence are not stored
-    # (tagger.Tagger.infer). Default 0.70 — the suggestion path already
-    # filters at 0.70 and the centroid/learned path covers low-confidence
-    # preferred tags, so the sub-0.70 tail is redundant weight (it had
-    # bloated image_record's TOAST to ~100 GB; plan-task #764). Operator-
-    # tunable via Settings → ML; must stay ≤ the suggestion thresholds.
-    tagger_store_floor: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.70
-    )
-    min_reference_images: Mapped[int] = mapped_column(
-        Integer, nullable=False, default=5
-    )
-    # Video tagging (#747). Sample one frame every N seconds (fixed CADENCE, not a
-    # fixed count) so a tag's frame-presence reflects real screen time regardless
-    # of video length; cap the total so a long video can't explode into hundreds
-    # of inferences (the cadence stretches past the cap). A tag is kept only if it
-    # appears in >= video_min_tag_frames sampled frames (≈ that many × interval
-    # seconds on screen) — duration-independent noise rejection. Operator-tunable.
     video_frame_interval_seconds: Mapped[float] = mapped_column(
         Float, nullable=False, default=4.0
     )
     video_max_frames: Mapped[int] = mapped_column(
         Integer, nullable=False, default=64
     )
-    video_min_tag_frames: Mapped[int] = mapped_column(
-        Integer, nullable=False, default=3
-    )
     # Tagging-v2 head training (#114). The head is the suggestion source that
     # LEARNS from the operator's tags (replacing Camie + centroid). A concept
     # needs >= head_min_positives labelled images before a head is trained;
@@ -101,11 +80,16 @@ class MLSettings(Base):
     ccip_auto_apply_threshold: Mapped[float] = mapped_column(
         Float, nullable=False, default=0.92
     )
-    tagger_model_version: Mapped[str] = mapped_column(
+    # Default = SigLIP 2 (so400m, 512px) for new installs (migration 0069);
-        String(128), nullable=False, default="camie-tagger-v2"
+    # existing libraries keep their stored value until the operator re-embeds.
-    )
     embedder_model_version: Mapped[str] = mapped_column(
-        String(128), nullable=False, default="siglip-so400m-patch14-384"
+        String(128), nullable=False, default="siglip2-so400m-patch16-512"
+    )
+    # The HF model NAME the embedder loads (server CPU embed + announced to the
+    # GPU agent in the lease). Operator-settable so the embedder is a choice, not
+    # a hardcode (#1190): set name + version together, then re-embed + retrain.
+    embedder_model_name: Mapped[str] = mapped_column(
+        String(128), nullable=False, default="google/siglip2-so400m-patch16-512"
     )
     updated_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()

backend/app/models/tag_allowlist.py

@@ -1,32 +0,0 @@
-"""TagAllowlist — tags the operator opted-in to auto-apply via maintenance."""
-
-from datetime import datetime
-
-from sqlalchemy import CheckConstraint, DateTime, Float, ForeignKey, func
-from sqlalchemy.orm import Mapped, mapped_column
-
-from .base import Base
-
-
-class TagAllowlist(Base):
-    __tablename__ = "tag_allowlist"
-    # Bare name — Base.metadata's naming convention prepends ck_<table>_,
-    # producing the final ck_tag_allowlist_confidence_range (matches migration 0003).
-    __table_args__ = (
-        CheckConstraint(
-            "min_confidence > 0 AND min_confidence <= 1",
-            name="confidence_range",
-        ),
-    )
-
-    tag_id: Mapped[int] = mapped_column(
-        ForeignKey("tag.id", ondelete="CASCADE"), primary_key=True
-    )
-    # Default auto-apply threshold for a newly-accepted tag. 0.90 (lowered from
-    # 0.95 on operator evidence 2026-06-07: 0.95 was too strict and skipped
-    # confident-enough applications). Per-tag value is still tunable in the
-    # allowlist table; existing rows keep whatever they were stored with.
-    min_confidence: Mapped[float] = mapped_column(Float, nullable=False, default=0.90)
-    added_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now()
-    )

backend/app/models/tag_eval_run.py

@@ -1,45 +0,0 @@
-"""TagEvalRun — persisted lifecycle of a head-vs-centroid tagging eval (#1130).
-
-Mirrors LibraryAuditRun so the result SURVIVES navigation: the run + its full
-report live in this row, and the admin card rehydrates from it on mount instead
-of holding the report in transient frontend state. State machine:
-running → ready / error. The async ml-queue task writes `report` (JSONB) when
-done; a maintenance recovery sweep flips a stalled `running` row to `error`.
-"""
-
-from datetime import datetime
-from typing import Any
-
-from sqlalchemy import DateTime, Integer, String, Text, func
-from sqlalchemy.dialects.postgresql import JSONB
-from sqlalchemy.orm import Mapped, mapped_column
-
-from .base import Base
-
-
-class TagEvalRun(Base):
-    __tablename__ = "tag_eval_run"
-
-    id: Mapped[int] = mapped_column(Integer, primary_key=True)
-    # The eval parameters: {concepts: [...], curve_points: [...], neg_ratio,
-    # cv_folds, ...} — echoed back so the report is self-describing.
-    params: Mapped[dict[str, Any]] = mapped_column(JSONB, nullable=False)
-    status: Mapped[str] = mapped_column(
-        String(16), nullable=False, default="running", index=True,
-    )
-    # running | ready | error
-    started_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now(),
-    )
-    finished_at: Mapped[datetime | None] = mapped_column(
-        DateTime(timezone=True), nullable=True,
-    )
-    # The full result: per-concept metrics (head vs centroid), learning-curve
-    # points, and example image ids. Null until the task finishes.
-    report: Mapped[dict[str, Any] | None] = mapped_column(JSONB, nullable=True)
-    error: Mapped[str | None] = mapped_column(Text, nullable=True)
-    # Last time the task made progress — the recovery sweep tells a live run
-    # from a SIGKILL'd one by this (mirrors LibraryAuditRun).
-    last_progress_at: Mapped[datetime | None] = mapped_column(
-        DateTime(timezone=True), nullable=True,
-    )

backend/app/models/tag_reference_embedding.py

@@ -1,23 +0,0 @@
-"""TagReferenceEmbedding — per-tag centroid (mean SigLIP embedding of members)."""
-
-from datetime import datetime
-
-from pgvector.sqlalchemy import Vector
-from sqlalchemy import DateTime, ForeignKey, Integer, String, func
-from sqlalchemy.orm import Mapped, mapped_column
-
-from .base import Base
-
-
-class TagReferenceEmbedding(Base):
-    __tablename__ = "tag_reference_embedding"
-
-    tag_id: Mapped[int] = mapped_column(
-        ForeignKey("tag.id", ondelete="CASCADE"), primary_key=True
-    )
-    embedding: Mapped[list[float]] = mapped_column(Vector(1152), nullable=False)
-    reference_count: Mapped[int] = mapped_column(Integer, nullable=False)
-    model_version: Mapped[str] = mapped_column(String(128), nullable=False)
-    updated_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now()
-    )

backend/app/scripts/download_models.py

@@ -7,7 +7,6 @@ import sys
 from pathlib import Path
 
 MODEL_ROOT = Path(os.environ.get("ML_MODEL_DIR", "/models"))
-CAMIE_REPO = os.environ.get("CAMIE_HF_REPO", "Camais03/camie-tagger-v2")
 SIGLIP_REPO = os.environ.get(
     "SIGLIP_HF_REPO", "google/siglip-so400m-patch14-384"
 )
@@ -24,34 +23,6 @@ def _snapshot(repo_id: str, dest: Path, allow_patterns: list[str] | None) -> Non
     )
 
 
-def ensure_camie() -> None:
-    """Fetch Camie v2 weights + metadata.
-
-    v2 layout (HuggingFace Camais03/camie-tagger-v2): the ONNX file is
-    named camie-tagger-v2.onnx (not model.onnx) and tags ship inside
-    camie-tagger-v2-metadata.json (not selected_tags.csv). Both at root.
-    The repo also contains app/, game/, training/, images/ subdirs full
-    of setup/demo files we don't need — allow_patterns scopes the fetch
-    to just the inference essentials (~790 MB instead of ~2 GB).
-    """
-    dest = MODEL_ROOT / "camie"
-    model_file = dest / "camie-tagger-v2.onnx"
-    meta_file = dest / "camie-tagger-v2-metadata.json"
-    if model_file.is_file() and meta_file.is_file():
-        print(f"[download_models] Camie present at {dest}")
-        return
-    print(f"[download_models] Fetching {CAMIE_REPO} -> {dest}")
-    _snapshot(
-        CAMIE_REPO, dest,
-        [
-            "camie-tagger-v2.onnx",
-            "camie-tagger-v2-metadata.json",
-            "config.json",
-            "config.yaml",
-        ],
-    )
-
-
 def ensure_siglip() -> None:
     dest = MODEL_ROOT / "siglip"
     if (dest / "config.json").is_file() and any(dest.glob("*.safetensors")):
@@ -62,7 +33,6 @@ def ensure_siglip() -> None:
 
 
 def main() -> int:
-    ensure_camie()
     ensure_siglip()
     print("[download_models] Done.")
     return 0

backend/app/services/audits/single_color.py

@@ -1,8 +1,9 @@
 """Single-color audit: matches images where one color dominates beyond
-the threshold (within the given Euclidean RGB tolerance). The first
+the threshold (within the given Euclidean RGB tolerance). The canonical
-canonical implementation — the import-side filter (SkipReason.single_color)
+predicate for BOTH surfaces: FC-Cleanup's retroactive audit and — since
-was never wired; FC-Cleanup's audit module is the source of truth and a
+2026-07-02 — the import-side filter (Importer._single_color_hit /
-future spec can adopt it on the import path too.
+SkipReason.single_color), so what the audit flags and what the import
+skips can never disagree.
 """
 
 from PIL import Image

backend/app/services/cleanup_service.py

@@ -395,9 +395,8 @@ def delete_images(
 def delete_tag(session: Session, *, tag_id: int) -> dict:
     """Simple DELETE FROM tag WHERE id=?.
 
-    Postgres cascades the rest (image_tag, tag_alias, tag_allowlist,
+    Postgres cascades the rest (image_tag, tag_alias, tag_suggestion_rejection,
-    tag_reference_embedding, tag_suggestion_rejection, series_page).
+    series_page). Returns counts BEFORE delete so the caller can surface them.
-    Returns counts BEFORE delete so the caller can surface them.
     Raises LookupError if tag_id not found.
     """
     tag = session.get(Tag, tag_id)
@@ -719,89 +718,24 @@ def reconcile_duplicate_posts(
     return {"groups": len(groups), "merged": losers_total, "sample": sample}
 
 
-# Legacy tags FC no longer uses, in two shapes:
+# The CONTENT vocabulary. "Reset content tagging" wipes these so the operator
-#   (1) kinds the tag input never produces — archive/post/artist.
+# can re-tag from scratch. fandom + series (and series_page ordering) are
-#       provenance (post grouping) + archive membership are their own
+# deliberately NOT here — they're kept.
-#       systems now, and artists are first-class Artist/Source rows.
-#       meta/rating were already hard-deleted by alembic 0023.
-#   (2) name prefixes from IR kinds FC never adopted — `source:*`.
-#       ImageRepo had a `source` kind; FC's enum doesn't, so ir_ingest
-#       fell those back to `general` (kind=general, name="source:patreon"
-#       etc.). They can't be caught by kind, so we match the name prefix.
-PURGEABLE_TAG_KINDS = ("archive", "post", "artist")
-LEGACY_NAME_PREFIXES = ("source:",)
-
-
-def _legacy_tag_predicate():
-    name_clauses = [Tag.name.like(f"{p}%") for p in LEGACY_NAME_PREFIXES]
-    return or_(Tag.kind.in_(PURGEABLE_TAG_KINDS), *name_clauses)
-
-
-def purge_legacy_tags(session: Session, *, dry_run: bool = False) -> dict:
-    """Count (dry_run) or delete legacy IR-migration tags: archive/post/
-    artist-kind tags PLUS general tags whose name matches a legacy
-    prefix (source:*).
-
-    CASCADE on image_tag / tag_alias / tag_allowlist /
-    tag_reference_embedding / tag_suggestion_rejection / series_page
-    clears the related rows on the parent DELETE.
-
-    Returns:
-      {"by_kind": {kind: count, ...},        # kind-matched rows
-       "by_prefix": {"source:*": count},     # name-prefix-matched rows
-       "count": total, "sample_names": [first 50],
-       and on live runs "deleted": total}
-    """
-    predicate = _legacy_tag_predicate()
-    rows = session.execute(
-        select(Tag.id, Tag.name, Tag.kind).where(predicate)
-    ).all()
-    by_kind: dict[str, int] = {}
-    by_prefix: dict[str, int] = {}
-    for _id, name, kind in rows:
-        # Classify by name-prefix first so a source:* row counts once,
-        # under the prefix bucket, regardless of its (general) kind.
-        matched_prefix = next(
-            (p for p in LEGACY_NAME_PREFIXES if name.startswith(p)), None,
-        )
-        if matched_prefix is not None:
-            label = f"{matched_prefix}*"
-            by_prefix[label] = by_prefix.get(label, 0) + 1
-        else:
-            key = kind.value if hasattr(kind, "value") else str(kind)
-            by_kind[key] = by_kind.get(key, 0) + 1
-    sample = [name for _id, name, _kind in rows[:50]]
-    total = len(rows)
-    result = {
-        "by_kind": by_kind, "by_prefix": by_prefix,
-        "count": total, "sample_names": sample,
-    }
-    if dry_run:
-        return result
-    if total:
-        session.execute(Tag.__table__.delete().where(predicate))
-        session.commit()
-    result["deleted"] = total
-    return result
-
-
-# The Camie-suggestable CONTENT vocabulary. "Reset content tagging" wipes
-# these so the operator can re-tag from scratch via auto-suggest. fandom +
-# series (and series_page ordering) are deliberately NOT here — they're kept.
 RESETTABLE_TAG_KINDS = ("general", "character")
 
 
 def reset_content_tagging(session: Session, *, dry_run: bool = False) -> dict:
     """Count (dry_run) or DELETE every general + character tag so the operator
-    can re-tag from scratch via the Camie auto-suggest.
+    can re-tag from scratch. NB: the deleted applications include the tagging
+    heads' training positives — suggestions do NOT repopulate on their own; the
+    heads retrain from whatever the operator re-tags. (The API route gates the
+    live run behind a preview-derived confirm token for exactly this reason.)
 
-    PRESERVED: fandom + series tags and their series_page ordering, plus every
+    PRESERVED: fandom + series tags and their series_page ordering. CASCADE on
-    image's image_prediction rows (untouched) so suggestions
+    image_tag / tag_alias / tag_suggestion_rejection clears each deleted tag's
-    repopulate immediately. CASCADE on image_tag / tag_alias / tag_allowlist /
+    applications + metadata. Tag.fandom_id is SET NULL, so deleting character
-    tag_reference_embedding / tag_suggestion_rejection clears each deleted
+    tags never touches the fandom rows. Irreversible except via DB backup
-    tag's applications + metadata. Tag.fandom_id is SET NULL, so deleting
+    restore.
-    character tags never touches the fandom rows. Irreversible except via DB
-    backup restore.
 
     Returns:
       {"by_kind": {"general": N, "character": M},
@@ -1074,7 +1008,7 @@ def reextract_archive_attachments(
     still an archive on disk, so the cursor is what guarantees forward progress.
     """
     from ..models import ImportSettings, Post, PostAttachment, Source
-    from ..tasks.ml import tag_and_embed
+    from ..tasks.ml import cpu_embed_enabled, embed_image
     from ..tasks.thumbnail import generate_thumbnail
     from .archive_extractor import is_archive
     from .importer import Importer
@@ -1155,10 +1089,12 @@ def reextract_archive_attachments(
 
     # Thumbnails + ML for the newly-imported members (best-effort; off the
     # critical path — a Redis hiccup must not fail the whole re-extract).
+    do_embed = cpu_embed_enabled()
     for img_id in enqueue_ids:
         try:
             generate_thumbnail.delay(img_id)
-            tag_and_embed.delay(img_id)
+            if do_embed:
+                embed_image.delay(img_id)
         except Exception as exc:
             log.warning("re-extract enqueue failed for image %s: %s", img_id, exc)
     return summary

backend/app/services/download_service.py

@@ -326,14 +326,16 @@ class DownloadService:
                 # for hours after a download landed. Lazy import to avoid
                 # circular-import risk between this service and the
                 # tasks/* modules that import it.
-                from ..tasks.ml import tag_and_embed
+                from ..tasks.ml import cpu_embed_enabled, embed_image
                 from ..tasks.thumbnail import generate_thumbnail
+                do_embed = cpu_embed_enabled()
                 ids = list(result.member_image_ids)
                 if result.image_id is not None and result.image_id not in ids:
                     ids.append(result.image_id)
                 for img_id in ids:
                     generate_thumbnail.delay(img_id)
-                    tag_and_embed.delay(img_id)
+                    if do_embed:
+                        embed_image.delay(img_id)
             elif result.status == "attached":
                 # Non-media or extracted archive captured as PostAttachment
                 # (FC-2d-iii). The canonical copy lives in the attachments

backend/app/services/gallery_service.py

@@ -55,6 +55,25 @@ def _effective_date_col():
     return ImageRecord.effective_date
 
 
+# Sort key -> the materialized column the gallery orders + cursors on. Both are
+# indexed (DESC, id DESC), so every sort is a forward index range scan.
+#   newest/oldest    → effective_date    (primary post's date, else download)
+#   posted_new/_old  → earliest_post_date (earliest publish across ALL posts)
+_SORT_COLUMNS = {
+    "newest": ImageRecord.effective_date,
+    "oldest": ImageRecord.effective_date,
+    "posted_new": ImageRecord.earliest_post_date,
+    "posted_old": ImageRecord.earliest_post_date,
+}
+_ASCENDING_SORTS = {"oldest", "posted_old"}
+
+
+def _sort_column(sort: str):
+    """The materialized date column a gallery sort orders/cursors on (falls back
+    to effective_date for any unknown sort)."""
+    return _SORT_COLUMNS.get(sort, ImageRecord.effective_date)
+
+
 def _outer_join_primary_post(stmt: Select) -> Select:
     """LEFT JOIN Post on ImageRecord.primary_post_id so the COALESCE
     above sees Post.post_date when available. Images without a post
@@ -289,6 +308,80 @@ def _gallery_images(rows, artists: dict[int, dict]) -> list[GalleryImage]:
     ]
 
 
+def _diversify_similar(src, rows, limit, *, dup_threshold=8, lam=0.40):
+    """Trim a nearest-cosine candidate pool down to `limit` diverse picks.
+
+    1. pHash collapse: drop any candidate whose perceptual hash is within
+       `dup_threshold` Hamming bits of the anchor or an already-kept candidate —
+       so a reposted banner (and the anchor's own clones) appears at most once.
+    2. MMR (Maximal Marginal Relevance): greedily pick the candidate maximising
+       `lam * sim_to_anchor - (1 - lam) * max_sim_to_already_picked`. This keeps
+       the most relevant up top but pushes the selection to SPAN clusters
+       instead of returning 40 variations of one image.
+
+    `lam` is the variance dial: lower = weight the diversity penalty harder, so
+    the rail reaches further across clusters (operator wanted MORE variance,
+    2026-07-01 — dropped 0.55→0.40, dup 6→8, paired with a wider pool in
+    `similar()`).
+
+    Falls back to nearest-order (`rows[:limit]`) on any failure or a small pool.
+    """
+    if len(rows) <= 1:
+        return rows[:limit]
+    try:
+        import imagehash
+        import numpy as np
+    except Exception:
+        return rows[:limit]
+
+    # --- 1. pHash near-duplicate collapse (videos/NULL phash pass through) ---
+    kept = []
+    seen = []
+    if src.phash:
+        try:
+            seen.append(imagehash.hex_to_hash(src.phash))
+        except Exception:
+            pass
+    for row in rows:
+        ph = row[0].phash
+        if ph:
+            try:
+                h = imagehash.hex_to_hash(ph)
+                if any((h - k) <= dup_threshold for k in seen):
+                    continue
+                seen.append(h)
+            except Exception:
+                pass
+        kept.append(row)
+    if len(kept) <= limit:
+        return kept
+
+    # --- 2. MMR re-rank on the L2-normalised SigLIP embeddings ---
+    try:
+        a = np.asarray(src.siglip_embedding, dtype=np.float32)
+        a = a / (np.linalg.norm(a) or 1.0)
+        V = np.vstack([
+            np.asarray(row[0].siglip_embedding, dtype=np.float32) for row in kept
+        ])
+        V = V / np.clip(np.linalg.norm(V, axis=1, keepdims=True), 1e-8, None)
+    except Exception:
+        return kept[:limit]
+
+    rel = V @ a                                  # (N,) cosine to the anchor
+    n = len(kept)
+    picked_mask = np.zeros(n, dtype=bool)
+    max_sim = np.zeros(n, dtype=np.float32)      # max sim to anything picked yet
+    order = []
+    for _ in range(min(limit, n)):
+        scores = lam * rel - (1.0 - lam) * max_sim
+        scores[picked_mask] = -np.inf
+        i = int(np.argmax(scores))
+        order.append(i)
+        picked_mask[i] = True
+        max_sim = np.maximum(max_sim, V @ V[i])
+    return [kept[i] for i in order]
+
+
 async def _artists_for(session, image_ids: list[int]) -> dict[int, dict]:
     """Map image_id -> {"name","slug"} via the canonical
     image_record.artist_id (FC-2d-vii-c). Bounded by page size."""
@@ -332,7 +425,10 @@ class GalleryService:
             tag_ids, post_id, artist_id, tag_or_groups, tag_exclude,
         )
 
-        eff = _effective_date_col()
+        # eff is the ACTIVE sort column (effective_date or earliest_post_date);
+        # the cursor, ordering and year/month grouping all key off it, so the
+        # 'post date' sort paginates + buckets by original publish transparently.
+        eff = _sort_column(sort)
         stmt = select(ImageRecord, Post.post_date, eff.label("eff"))
         stmt = _outer_join_primary_post(stmt)
         stmt = _apply_scope(
@@ -343,7 +439,7 @@ class GalleryService:
             date_from=date_from, date_to=date_to,
         )
 
-        descending = sort != "oldest"
+        descending = sort not in _ASCENDING_SORTS
         if cursor:
             cur_ts, cur_id = decode_cursor(cursor)
             # The cursor is just (last eff, last id); the request's sort
@@ -565,14 +661,20 @@ class GalleryService:
         untagged: bool = False, no_artist: bool = False,
         date_from: datetime | None = None, date_to: datetime | None = None,
     ) -> list[GalleryImage] | None:
-        """Visual "more like this": images ranked by cosine distance to
+        """Visual "more like this": images near `image_id`'s SigLIP embedding
-        `image_id`'s SigLIP embedding (pgvector, HNSW-indexed — alembic 0036).
+        (pgvector, HNSW-indexed — alembic 0036), then DIVERSIFIED so the result
-        No ML inference here; the embedding was computed at import.
+        doesn't collapse into one cluster. No ML inference here.
 
-        Returns None if the source image doesn't exist (→ 404), [] if it has
+        Pure nearest-cosine piles up near-identical images — a reposted banner
-        no embedding (a video / not-yet-embedded). Composes with the Phase-1/2
+        fills the whole grid, and once you wander into a B&W / comic-panel
-        scope filters (AND) but REPLACES the date sort — always nearest-first,
+        cluster every neighbour is more of the same with no way back to colour
-        bounded to `limit` (no cursor; distance-ranking has no date cursor).
+        (operator-reported 2026-06-30). So we pull a WIDER candidate pool, then:
+          1. collapse near-duplicate pHashes (and drop clones of the anchor),
+          2. MMR re-rank — pick for closeness-to-anchor but penalise similarity
+             to what's already picked, so the result SPANS clusters.
+
+        Returns None if the source doesn't exist (→ 404), [] if it has no
+        embedding. Composes with the scope filters (AND); REPLACES the date sort.
         """
         if limit < 1 or limit > 200:
             raise ValueError("limit must be between 1 and 200")
@@ -582,6 +684,11 @@ class GalleryService:
         if src.siglip_embedding is None:
             return []
 
+        # Over-fetch so diversification has clusters to spread across — without a
+        # wide pool there's nothing but the near-dupes to choose from. Widened
+        # (5×→8×, cap 200→400) so the stronger MMR has genuinely distinct
+        # neighbourhoods to reach into for more variance (operator, 2026-07-01).
+        pool_n = min(400, max(limit * 8, 100))
         distance = ImageRecord.siglip_embedding.cosine_distance(src.siglip_embedding)
         eff = _effective_date_col()
         stmt = select(ImageRecord, Post.post_date, eff.label("eff"))
@@ -597,8 +704,9 @@ class GalleryService:
             platform=platform, untagged=untagged, no_artist=no_artist,
             date_from=date_from, date_to=date_to,
         )
-        stmt = stmt.order_by(distance.asc()).limit(limit)
+        stmt = stmt.order_by(distance.asc()).limit(pool_n)
         rows = (await self.session.execute(stmt)).all()
+        rows = _diversify_similar(src, rows, limit)
         artists = await _artists_for(self.session, [r[0].id for r in rows])
         return _gallery_images(rows, artists)
 

backend/app/services/importer.py

@@ -17,7 +17,7 @@ from enum import StrEnum
 from pathlib import Path
 
 from PIL import Image
-from sqlalchemy import select, update
+from sqlalchemy import func, select, update
 from sqlalchemy.exc import IntegrityError
 from sqlalchemy.orm import Session
 
@@ -44,6 +44,7 @@ from ..utils.sidecar import find_sidecar, parse_sidecar
 from ..utils.slug import slugify
 from .archive_extractor import extract_archive, is_archive
 from .attachment_store import AttachmentStore
+from .audits import single_color
 from .link_extract import extract_external_links
 from .thumbnailer import Thumbnailer
 
@@ -790,6 +791,13 @@ class Importer:
                         error=f"{pct:.2%} transparent",
                     )
 
+            if self.settings.skip_single_color and self._single_color_hit(source):
+                return ImportResult(
+                    status="skipped", skip_reason=SkipReason.single_color,
+                    error=(f"one color dominates >"
+                           f"{self.settings.single_color_threshold:.0%}"),
+                )
+
         # Artist anchored to the attribution path (folder→artist), resolved
         # UP-FRONT so the enrich-on-duplicate branches link provenance with the
         # right artist even when the sidecar carries none — which is now the norm
@@ -1123,6 +1131,13 @@ class Importer:
                         status="skipped", skip_reason=SkipReason.too_transparent,
                         error=f"{pct:.2%} transparent",
                     )
+
+            if self.settings.skip_single_color and self._single_color_hit(path):
+                return ImportResult(
+                    status="skipped", skip_reason=SkipReason.single_color,
+                    error=(f"one color dominates >"
+                           f"{self.settings.single_color_threshold:.0%}"),
+                )
         else:
             # Best-effort probe for dims + duration so downloaded videos can dedup
             # (#871). LENIENT: unlike _import_media this path does not reject on a
@@ -1411,6 +1426,24 @@ class Importer:
         # default (matches the old COALESCE(post_date, created_at) fallback).
         if record.primary_post_id == post.id and post.post_date is not None:
             record.effective_date = post.post_date
+        # earliest_post_date (alembic 0071) = MIN(post_date) across ALL of this
+        # image's provenance posts, not just the primary — so the gallery can
+        # sort by original publish rather than the download/repost the primary
+        # points at. Recompute from provenance whenever a dated post is linked;
+        # the provenance row for THIS post was committed above, so the MIN
+        # includes it. Leaves the created_at default when no linked post is dated.
+        if post.post_date is not None:
+            earliest = self.session.execute(
+                select(func.min(Post.post_date))
+                .select_from(ImageProvenance)
+                .join(Post, Post.id == ImageProvenance.post_id)
+                .where(
+                    ImageProvenance.image_record_id == record.id,
+                    Post.post_date.is_not(None),
+                )
+            ).scalar_one_or_none()
+            if earliest is not None:
+                record.earliest_post_date = earliest
         self.session.flush()
 
     def _copy_to_library(
@@ -1475,20 +1508,8 @@ class Importer:
         existing.duration_seconds = duration  # #871: keep the kept copy's duration
         existing.thumbnail_path = None
         existing.integrity_status = "unknown"
-        existing.tagger_model_version = None
         existing.siglip_embedding = None
         existing.siglip_model_version = None
-        existing.centroid_scores = None
-        # #768: predictions also live in the normalized image_prediction table
-        # now — clear them so a re-imported file re-derives a fresh set.
-        from sqlalchemy import delete as _delete
-
-        from ..models import ImagePrediction as _ImagePrediction
-        self.session.execute(
-            _delete(_ImagePrediction).where(
-                _ImagePrediction.image_record_id == existing.id
-            )
-        )
         # created_at intentionally preserved; updated_at auto-bumps.
         self.session.flush()
         self.session.commit()
@@ -1532,6 +1553,33 @@ class Importer:
                 # Benign orphan; the DB swap already committed. Don't undo it.
                 pass
 
+    # Matches the Cleanup audit card's default tolerance: the import-side
+    # filter and the retroactive audit must agree on what "single color" MEANS
+    # (Euclidean RGB distance to the dominant color); only the match threshold
+    # is operator-tunable per surface.
+    _SINGLE_COLOR_TOLERANCE = 30
+
+    def _single_color_hit(self, source: Path) -> bool:
+        """True when one color dominates beyond the configured threshold — the
+        same canonical predicate the Cleanup audit runs (audits.single_color,
+        whose docstring anticipated this adoption; the skip_single_color
+        setting existed but was never wired until 2026-07-02). Never raises:
+        unreadable files were already rejected by verify() upstream, and a
+        residual decode error just declines to match (the import proceeds)."""
+        try:
+            with Image.open(source) as im:
+                if getattr(im, "is_animated", False):
+                    # Frame 0 only would misjudge animations; skip like the
+                    # transparency check does.
+                    return False
+                return single_color.evaluate(
+                    im,
+                    threshold=self.settings.single_color_threshold,
+                    tolerance=self._SINGLE_COLOR_TOLERANCE,
+                )
+        except Exception:
+            return False
+
     def _transparency_pct(self, source: Path) -> float:
         """Fraction of fully-transparent pixels in the image. 0.0 if no alpha.
 

backend/app/services/ml/__init__.py

@@ -1 +1,3 @@
-"""ML pipeline services: tagger, embedder, suggestions, centroids, allowlist, aliases."""
+"""ML pipeline services: embedders, heads (the learning suggester), suggestions,
+GPU-job queue + failure triage, CCIP characters, crops/regions, allowlist and
+aliases."""

backend/app/services/ml/allowlist.py

@@ -1,36 +1,20 @@
-"""Allowlist semantics: accepting a suggestion adds the canonical tag to
+"""Suggestion actions: accept applies the canonical tag to an image (which
-image_tag AND to tag_allowlist; per-image removal/dismiss writes a rejection.
+feeds head training); dismiss / reject record a per-image rejection.
+
+(The Camie allowlist bulk-apply was retired #1189 — heads + CCIP are the tag
+source, and head auto-apply is the earned propagation. Accept no longer
+allowlists or fans a tag out across the library.)
 """
 
-from collections.abc import Sequence
+from sqlalchemy import delete
-from dataclasses import dataclass
-
-from sqlalchemy import and_, delete, distinct, func, or_, select
 from sqlalchemy.dialects.postgresql import insert
 from sqlalchemy.ext.asyncio import AsyncSession
 
-from ...models import (
+from ...models import TagSuggestionRejection
-    ImagePrediction,
-    MLSettings,
-    Tag,
-    TagAlias,
-    TagAllowlist,
-    TagSuggestionRejection,
-)
 from ...models.tag import image_tag
 from .aliases import AliasService
 
 
-@dataclass(frozen=True)
-class AllowlistRow:
-    tag_id: int
-    tag_name: str
-    tag_kind: str
-    min_confidence: float
-    applied_count: int      # image_tag rows currently carrying this tag
-    coverage_count: int     # images a sweep WOULD cover at min_confidence
-
-
 class AllowlistService:
     def __init__(self, session: AsyncSession):
         self.session = session
@@ -39,21 +23,11 @@ class AllowlistService:
     async def _apply_image_tag(self, image_id: int, tag_id: int, source: str):
         stmt = insert(image_tag).values(
             image_record_id=image_id, tag_id=tag_id, source=source
-        )
+        ).on_conflict_do_nothing(
-        stmt = stmt.on_conflict_do_nothing(
             index_elements=["image_record_id", "tag_id"]
         )
         await self.session.execute(stmt)
 
-    async def _add_to_allowlist(self, tag_id: int) -> bool:
-        """Returns True if newly added (caller should kick off retro-apply)."""
-        exists = await self.session.get(TagAllowlist, tag_id)
-        if exists is not None:
-            return False
-        self.session.add(TagAllowlist(tag_id=tag_id))
-        await self.session.flush()
-        return True
-
     async def _clear_rejection(self, image_id: int, tag_id: int):
         await self.session.execute(
             delete(TagSuggestionRejection)
@@ -61,12 +35,11 @@ class AllowlistService:
             .where(TagSuggestionRejection.tag_id == tag_id)
         )
 
-    async def accept(self, image_id: int, tag_id: int) -> bool:
+    async def accept(self, image_id: int, tag_id: int) -> None:
-        """Accept a suggestion. Returns True if the tag was newly added to
+        """Apply the accepted tag to this image (source='ml_accepted', a head
-        the allowlist (the API layer enqueues apply_allowlist_tags then)."""
+        training positive) and clear any prior rejection."""
         await self._apply_image_tag(image_id, tag_id, source="ml_accepted")
         await self._clear_rejection(image_id, tag_id)
-        return await self._add_to_allowlist(tag_id)
 
     async def add_alias_and_accept(
         self,
@@ -74,17 +47,16 @@ class AllowlistService:
         alias_string: str,
         alias_category: str,
         canonical_tag_id: int,
-    ) -> bool:
+    ) -> None:
         await self.aliases.create(
             alias_string, alias_category, canonical_tag_id
         )
-        return await self.accept(image_id, canonical_tag_id)
+        await self.accept(image_id, canonical_tag_id)
 
     async def dismiss(self, image_id: int, tag_id: int) -> None:
         stmt = insert(TagSuggestionRejection).values(
             image_record_id=image_id, tag_id=tag_id
-        )
+        ).on_conflict_do_nothing(
-        stmt = stmt.on_conflict_do_nothing(
             index_elements=["image_record_id", "tag_id"]
         )
         await self.session.execute(stmt)
@@ -96,118 +68,11 @@ class AllowlistService:
         await self._clear_rejection(image_id, tag_id)
 
     async def reject_applied_tag(self, image_id: int, tag_id: int) -> None:
-        """Operator removed an applied tag from an image. Remove the
+        """Operator removed an applied tag from an image. Remove the image_tag
-        image_tag row AND record a rejection so the allowlist won't
+        row AND record a rejection so head auto-apply won't re-apply it."""
-        re-apply it on the next maintenance sweep."""
         await self.session.execute(
             image_tag.delete()
             .where(image_tag.c.image_record_id == image_id)
             .where(image_tag.c.tag_id == tag_id)
         )
         await self.dismiss(image_id, tag_id)
-
-    async def _store_floor(self) -> float:
-        return (
-            await self.session.execute(
-                select(MLSettings.tagger_store_floor).where(MLSettings.id == 1)
-            )
-        ).scalar_one()
-
-    async def update_threshold(
-        self, tag_id: int, min_confidence: float
-    ) -> None:
-        row = await self.session.get(TagAllowlist, tag_id)
-        if row is not None:
-            # An allowlist tag can't auto-apply more permissively than the
-            # ingest store floor — predictions below tagger_store_floor aren't
-            # stored, so a lower min_confidence would behave identically to the
-            # floor. Clamp so the stored threshold matches actual behavior
-            # (#764).
-            floor = await self._store_floor()
-            row.min_confidence = max(min_confidence, floor)
-
-    async def remove(self, tag_id: int) -> None:
-        await self.session.execute(
-            delete(TagAllowlist).where(TagAllowlist.tag_id == tag_id)
-        )
-
-    async def _coverage_match(self, tag: Tag):
-        """The predicate over image_prediction rows that resolve to `tag`,
-        mirroring tasks.ml._confidence_for_tag's resolution: a prediction whose
-        raw_name equals the tag name (any category), OR an alias maps
-        (raw_name, category) -> this tag. Returns a SQLAlchemy boolean clause.
-        """
-        alias_rows = (
-            await self.session.execute(
-                select(TagAlias.alias_string, TagAlias.alias_category).where(
-                    TagAlias.canonical_tag_id == tag.id
-                )
-            )
-        ).all()
-        name_clause = ImagePrediction.raw_name == tag.name
-        alias_clauses = [
-            and_(
-                ImagePrediction.raw_name == a,
-                ImagePrediction.category == c,
-            )
-            for a, c in alias_rows
-        ]
-        return or_(name_clause, *alias_clauses) if alias_clauses else name_clause
-
-    async def coverage(self, tag_id: int, threshold: float) -> int:
-        """How many distinct images a sweep WOULD cover for this tag at
-        `threshold`: images with a resolving prediction scoring >= threshold.
-        The gross candidate pool (NOT minus already-applied/rejected) — it's
-        the tuning signal for "lower the threshold and ~N more images qualify".
-        """
-        tag = await self.session.get(Tag, tag_id)
-        if tag is None:
-            return 0
-        match = await self._coverage_match(tag)
-        stmt = select(
-            func.count(distinct(ImagePrediction.image_record_id))
-        ).where(ImagePrediction.score >= threshold, match)
-        return (await self.session.execute(stmt)).scalar_one()
-
-    async def list_all(self) -> Sequence[AllowlistRow]:
-        stmt = (
-            select(
-                TagAllowlist.tag_id,
-                Tag.name,
-                Tag.kind,
-                TagAllowlist.min_confidence,
-            )
-            .join(Tag, Tag.id == TagAllowlist.tag_id)
-            .order_by(Tag.name.asc())
-        )
-        rows = (await self.session.execute(stmt)).all()
-        tag_ids = [r[0] for r in rows]
-
-        # Applied counts in ONE grouped query (vs N per-row counts).
-        applied: dict[int, int] = {}
-        if tag_ids:
-            applied = dict(
-                (
-                    await self.session.execute(
-                        select(image_tag.c.tag_id, func.count())
-                        .where(image_tag.c.tag_id.in_(tag_ids))
-                        .group_by(image_tag.c.tag_id)
-                    )
-                ).all()
-            )
-
-        result = []
-        for r in rows:
-            # Coverage is per-tag (alias set differs); allowlist is small.
-            cov = await self.coverage(r[0], r[3])
-            result.append(
-                AllowlistRow(
-                    tag_id=r[0],
-                    tag_name=r[1],
-                    tag_kind=r[2].value if hasattr(r[2], "value") else str(r[2]),
-                    min_confidence=r[3],
-                    applied_count=applied.get(r[0], 0),
-                    coverage_count=cov,
-                )
-            )
-        return result

backend/app/services/ml/centroids.py

@@ -1,163 +0,0 @@
-"""Tag centroids: the mean SigLIP embedding of a tag's member images.
-
-Powers centroid-augmented suggestions (a tag whose centroid is close to an
-image's embedding becomes a suggestion even if Camie didn't predict it).
-"""
-
-from dataclasses import dataclass
-
-import numpy as np
-from sqlalchemy import func, select
-from sqlalchemy.dialects.postgresql import insert
-from sqlalchemy.ext.asyncio import AsyncSession
-
-from ...models import (
-    ImageRecord,
-    MLSettings,
-    Tag,
-    TagKind,
-    TagReferenceEmbedding,
-)
-from ...models.tag import image_tag
-
-ELIGIBLE_KINDS = {
-    TagKind.character,
-    TagKind.fandom,
-    TagKind.general,
-    TagKind.series,
-}
-
-
-@dataclass(frozen=True)
-class CentroidHit:
-    tag_id: int
-    similarity: float
-
-
-class CentroidService:
-    def __init__(self, session: AsyncSession):
-        self.session = session
-
-    async def _min_reference_images(self) -> int:
-        return (
-            await self.session.execute(
-                select(MLSettings.min_reference_images).where(MLSettings.id == 1)
-            )
-        ).scalar_one()
-
-    async def _model_version(self) -> str:
-        """Audit 2026-06-02: SigLIP model-version stamp comes from the
-        DB row, not the env constant. tag_and_embed (tasks/ml.py:110)
-        already reads from MLSettings.embedder_model_version, so by
-        sourcing centroid stamps + drift checks from the same row, we
-        eliminate the silent-drift case the audit flagged. env
-        SIGLIP_MODEL_VERSION still drives which model embedder.py
-        loads at runtime; the version stamp is purely the operator-
-        controlled identifier."""
-        return (
-            await self.session.execute(
-                select(MLSettings.embedder_model_version).where(MLSettings.id == 1)
-            )
-        ).scalar_one()
-
-    async def recompute_for_tag(self, tag_id: int) -> bool:
-        """Recompute one tag's centroid. Returns True if a centroid was
-        written, False if skipped (ineligible kind or too few members)."""
-        tag = await self.session.get(Tag, tag_id)
-        if tag is None or tag.kind not in ELIGIBLE_KINDS:
-            return False
-
-        min_refs = await self._min_reference_images()
-
-        stmt = (
-            select(ImageRecord.siglip_embedding)
-            .join(image_tag, image_tag.c.image_record_id == ImageRecord.id)
-            .where(image_tag.c.tag_id == tag_id)
-            .where(ImageRecord.siglip_embedding.is_not(None))
-        )
-        embeddings = [
-            np.array(e, dtype=np.float32)
-            for e in (await self.session.execute(stmt)).scalars().all()
-        ]
-        if len(embeddings) < min_refs:
-            return False
-
-        centroid = np.mean(np.stack(embeddings), axis=0).astype(np.float32)
-        model_version = await self._model_version()
-
-        stmt = insert(TagReferenceEmbedding).values(
-            tag_id=tag_id,
-            embedding=centroid.tolist(),
-            reference_count=len(embeddings),
-            model_version=model_version,
-        )
-        stmt = stmt.on_conflict_do_update(
-            index_elements=["tag_id"],
-            set_={
-                "embedding": centroid.tolist(),
-                "reference_count": len(embeddings),
-                "model_version": model_version,
-                "updated_at": func.now(),
-            },
-        )
-        await self.session.execute(stmt)
-        return True
-
-    async def list_drifted(self) -> list[int]:
-        """Tag ids whose centroid is stale: member count != reference_count,
-        OR no centroid row, OR centroid built on a different SigLIP version.
-        Only considers eligible-kind tags with embeddings present."""
-        current_model_version = await self._model_version()
-        member_counts = (
-            select(
-                image_tag.c.tag_id.label("tag_id"),
-                func.count(image_tag.c.image_record_id).label("members"),
-            )
-            .join(ImageRecord, ImageRecord.id == image_tag.c.image_record_id)
-            .where(ImageRecord.siglip_embedding.is_not(None))
-            .group_by(image_tag.c.tag_id)
-            .subquery()
-        )
-        stmt = (
-            select(Tag.id)
-            .join(member_counts, member_counts.c.tag_id == Tag.id)
-            .outerjoin(
-                TagReferenceEmbedding,
-                TagReferenceEmbedding.tag_id == Tag.id,
-            )
-            .where(Tag.kind.in_(ELIGIBLE_KINDS))
-            .where(
-                (TagReferenceEmbedding.tag_id.is_(None))
-                | (
-                    TagReferenceEmbedding.reference_count
-                    != member_counts.c.members
-                )
-                | (TagReferenceEmbedding.model_version != current_model_version)
-            )
-        )
-        return list((await self.session.execute(stmt)).scalars().all())
-
-    async def find_similar_tags(
-        self, image_id: int, limit: int = 20
-    ) -> list[CentroidHit]:
-        """Cosine similarity between an image's embedding and stored
-        centroids. Returns top-`limit` by similarity DESC. pgvector's
-        cosine_distance gives 1 - cosine_similarity."""
-        img = await self.session.get(ImageRecord, image_id)
-        if img is None or img.siglip_embedding is None:
-            return []
-        emb = img.siglip_embedding
-        distance = TagReferenceEmbedding.embedding.cosine_distance(emb)
-        stmt = (
-            select(
-                TagReferenceEmbedding.tag_id,
-                (1 - distance).label("similarity"),
-            )
-            .order_by(distance.asc())
-            .limit(limit)
-        )
-        rows = (await self.session.execute(stmt)).all()
-        return [
-            CentroidHit(tag_id=r.tag_id, similarity=float(r.similarity))
-            for r in rows
-        ]

backend/app/services/ml/embedder.py

@@ -18,9 +18,11 @@ ImageFile.LOAD_TRUNCATED_IMAGES = True
 # N_replicas × this within the cores allotted to ML to avoid oversubscription.
 _INTRA_OP_THREADS = 4
 
-MODEL_NAME = os.environ.get(
+DEFAULT_MODEL_NAME = os.environ.get(
     "SIGLIP_MODEL_NAME", "google/siglip-so400m-patch14-384"
 )
+# Back-compat alias (api/gpu imported this name as the fallback embedder id).
+MODEL_NAME = DEFAULT_MODEL_NAME
 MODEL_VERSION = os.environ.get(
     "SIGLIP_MODEL_VERSION", "siglip-so400m-patch14-384"
 )
@@ -29,35 +31,42 @@ _LOCAL_DIR = Path(os.environ.get("ML_MODEL_DIR", "/models")) / "siglip"
 
 
 class Embedder:
-    def __init__(self, model_dir: Path | None = None):
+    """Loads whatever SigLIP-family model it's given by HF NAME. For the default
-        self._model_dir = model_dir or _LOCAL_DIR
+    model it prefers the pre-downloaded local dir (no re-download on existing
+    deploys); any other name resolves as an HF repo id (downloaded + cached on
+    first use), so an operator model swap (#1190) just works server-side."""
+
+    def __init__(self, model_name: str | None = None, model_dir: Path | None = None):
+        self.model_name = model_name or DEFAULT_MODEL_NAME
+        self._explicit_dir = model_dir
         self._model = None
         self._processor = None
         self._torch = None
 
+    def _source(self) -> str:
+        if self._explicit_dir is not None:
+            return str(self._explicit_dir)
+        if self.model_name == DEFAULT_MODEL_NAME and _LOCAL_DIR.exists():
+            return str(_LOCAL_DIR)
+        return self.model_name
+
     def load(self) -> None:
         if self._model is not None:
             return
         import torch
-        from transformers import AutoModel, SiglipImageProcessor
+        from transformers import AutoImageProcessor, AutoModel
 
         self._torch = torch
         # Bound torch's CPU thread pool (see _INTRA_OP_THREADS) so each replica
         # stays a predictable core consumer on a shared node.
         torch.set_num_threads(_INTRA_OP_THREADS)
-        # FC's embedder only does IMAGE inference — never text. AutoProcessor
+        # IMAGE inference only — AutoImageProcessor loads just the image side
-        # loads the full processor including SiglipTokenizer, which requires
+        # (preprocessor_config.json), skipping the SigLIP tokenizer + its
-        # the sentencepiece library at import time even if we never call it.
+        # sentencepiece dep (operator hit that ImportError 2026-05-25). Works
-        # SiglipImageProcessor loads ONLY preprocessor_config.json (image
+        # for any SigLIP-family model, keeping the embedder model-agnostic.
-        # side) and skips the tokenizer config entirely. Operator hit the
+        src = self._source()
-        # ImportError 2026-05-25 once the ml-worker started actually running
+        self._processor = AutoImageProcessor.from_pretrained(src)
-        # tag_and_embed; switching to the image-only loader avoids the
+        self._model = AutoModel.from_pretrained(src)
-        # tokenizer dep without adding ~30 MB of unused C++ build to the
-        # lean ml-worker image.
-        self._processor = SiglipImageProcessor.from_pretrained(
-            str(self._model_dir)
-        )
-        self._model = AutoModel.from_pretrained(str(self._model_dir))
         self._model.eval()
 
     def infer(self, image_path: Path) -> np.ndarray:
@@ -74,8 +83,12 @@ class Embedder:
 _default_embedder: Embedder | None = None
 
 
-def get_embedder() -> Embedder:
+def get_embedder(model_name: str | None = None) -> Embedder:
+    """Cached embedder for `model_name` (default if None). Rebuilds the singleton
+    when the requested name changes, so an operator model swap takes effect
+    without restarting the worker."""
     global _default_embedder
-    if _default_embedder is None:
+    name = model_name or DEFAULT_MODEL_NAME
-        _default_embedder = Embedder()
+    if _default_embedder is None or _default_embedder.model_name != name:
+        _default_embedder = Embedder(model_name=name)
     return _default_embedder

backend/app/services/ml/gpu_jobs.py

@@ -12,8 +12,9 @@ and the lease itself reclaims expired leases as a final backstop. Result-writing
 
 from datetime import UTC, datetime, timedelta
 
-from sqlalchemy import and_, or_, select, update
+from sqlalchemy import and_, delete, exists, func, or_, select, update
 from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.orm import aliased
 
 from ...models import GpuJob
 
@@ -24,6 +25,107 @@ DEFAULT_LEASE_TTL = 180     # seconds an agent holds a job before it can be re-l
 DEFAULT_BATCH = 8
 MAX_ATTEMPTS = 3
 
+# Poison-loop backstops. `attempts` counts LEASES GRANTED (incremented in
+# lease()), but fail()'s MAX_ATTEMPTS cap only fires when the agent reports a
+# failure — a job that keeps coming back via release() (transient handback) or
+# lease expiry (agent crash/wedge) never gets a verdict and would cycle forever.
+# The orphan sweep converts those to 'error': an expired lease that has already
+# been granted EXPIRED_POISON_CAP leases is presumed to kill/wedge the agent,
+# and a pending job granted PENDING_POISON_CAP leases without ever completing is
+# presumed poisoned (e.g. a transfer that stalls every time). Both stay
+# resurrectable via /retry_errors, which resets attempts.
+EXPIRED_POISON_CAP = MAX_ATTEMPTS + 2
+PENDING_POISON_CAP = 10
+
+
+def error_dedupe_statements():
+    """DELETEs enforcing: at most ONE error row per (image, task), and none that
+    a live or succeeded row makes moot. The 2026-07-02 tombstone loop (backfill
+    skip-lists lacked 'error') minted a duplicate error row per bad file per
+    hour; running these before every backfill and inside /retry_errors keeps the
+    error count reading as "distinct failing files" and stops a retry fanning
+    one file out into several duplicate pending jobs. Shared by the sync beat
+    task and the async API route so both prune by the SAME predicate.
+    Execution order matters: moot rows first, then older duplicates (the newest
+    error — the freshest reason — survives)."""
+    other = aliased(GpuJob)
+    same_pair = and_(
+        other.image_record_id == GpuJob.image_record_id,
+        other.task == GpuJob.task,
+    )
+    moot = (
+        delete(GpuJob)
+        .where(
+            GpuJob.status == "error",
+            exists().where(
+                same_pair, other.status.in_(["pending", "leased", "done"]),
+            ),
+        )
+        .execution_options(synchronize_session=False)
+    )
+    older_dupe = (
+        delete(GpuJob)
+        .where(
+            GpuJob.status == "error",
+            exists().where(
+                same_pair,
+                other.status == "error",
+                or_(
+                    other.updated_at > GpuJob.updated_at,
+                    and_(other.updated_at == GpuJob.updated_at,
+                         other.id > GpuJob.id),
+                ),
+            ),
+        )
+        .execution_options(synchronize_session=False)
+    )
+    return [moot, older_dupe]
+
+
+def recover_statements(now: datetime) -> dict:
+    """UPDATEs for the orphan sweep, keyed by outcome; insertion order IS the
+    required execution order ('recovered' must run after 'poison_expired', which
+    claims the crash-loopers out of the same expired-lease pool)."""
+    expired = and_(GpuJob.status == "leased", GpuJob.lease_expires_at < now)
+    unlease = {"lease_token": None, "leased_at": None, "lease_expires_at": None,
+               "updated_at": now}
+    return {
+        "poison_expired": (
+            update(GpuJob)
+            .where(expired, GpuJob.attempts >= EXPIRED_POISON_CAP)
+            .values(
+                status="error",
+                # Keep the job's last stored failure reason — it's the triage
+                # signal for WHY the loop happened.
+                error=func.concat(
+                    f"poisoned: lease expired after {EXPIRED_POISON_CAP}+ lease "
+                    "attempts (job repeatedly crashes or wedges the agent?); "
+                    "last error: ",
+                    func.coalesce(GpuJob.error, "none"),
+                ),
+                **unlease,
+            )
+        ),
+        "recovered": update(GpuJob).where(expired).values(
+            status="pending", **unlease,
+        ),
+        "poison_pending": (
+            update(GpuJob)
+            .where(GpuJob.status == "pending",
+                   GpuJob.attempts >= PENDING_POISON_CAP)
+            .values(
+                status="error",
+                error=func.concat(
+                    f"poisoned: {PENDING_POISON_CAP}+ lease attempts without "
+                    "ever completing (transfer stalls every time?); "
+                    "last error: ",
+                    func.coalesce(GpuJob.error, "none"),
+                ),
+                updated_at=now,
+            )
+        ),
+    }
+
 
 class GpuJobService:
     def __init__(self, session: AsyncSession):
@@ -51,25 +153,33 @@ class GpuJobService:
     async def lease(
         self, token: str, batch_size: int = DEFAULT_BATCH, ttl: int = DEFAULT_LEASE_TTL
     ) -> list[GpuJob]:
-        """Claim up to batch_size pending (or expired-leased) jobs for `token`."""
+        """Claim up to batch_size pending (or expired-leased) jobs for `token`.
+
+        Two phases so each hits a partial index (0070) and stays O(batch) no
+        matter how many done/error rows have accumulated: the pending pool is the
+        hot path; expired leases are reclaimed only when pending can't fill the
+        batch (a crashed agent's work — rare). The old single OR-query walked the
+        primary key past the whole done-prefix in id order → O(done), which is
+        why leasing crawled — and the DB saturated — as the run progressed."""
         now = datetime.now(UTC)
-        picked = (
+
-            await self.session.execute(
+        async def _claim(condition, limit: int) -> list[int]:
-                select(GpuJob.id)
+            return list(
-                .where(
+                (
-                    or_(
+                    await self.session.execute(
-                        GpuJob.status == "pending",
+                        select(GpuJob.id).where(condition)
-                        and_(
+                        .order_by(GpuJob.id).limit(limit)
-                            GpuJob.status == "leased",
+                        .with_for_update(skip_locked=True)
-                            GpuJob.lease_expires_at < now,
-                        ),
                     )
-                )
+                ).scalars().all()
-                .order_by(GpuJob.id)
+            )
-                .limit(batch_size)
+
-                .with_for_update(skip_locked=True)
+        picked = await _claim(GpuJob.status == "pending", batch_size)
+        if len(picked) < batch_size:     # pending exhausted → reclaim expired leases
+            picked += await _claim(
+                and_(GpuJob.status == "leased", GpuJob.lease_expires_at < now),
+                batch_size - len(picked),
             )
-        ).scalars().all()
         if not picked:
             return []
         await self.session.execute(
@@ -162,16 +272,11 @@ class GpuJobService:
 
     async def recover_orphaned(self) -> int:
         """Reset every expired lease back to pending — catches agents that died
-        mid-job (no graceful release). Run on a short beat so the queue recovers
+        mid-job (no graceful release) — and convert poison-loopers to 'error'
-        + reads honestly even when no worker is actively leasing. Returns rows
+        (see the *_POISON_CAP rationale above). Run on a short beat so the queue
-        recovered."""
+        recovers + reads honestly even when no worker is actively leasing.
-        now = datetime.now(UTC)
+        Returns rows recovered to pending (poison conversions are extra)."""
-        res = await self.session.execute(
+        counts = {}
-            update(GpuJob)
+        for name, stmt in recover_statements(datetime.now(UTC)).items():
-            .where(GpuJob.status == "leased", GpuJob.lease_expires_at < now)
+            counts[name] = (await self.session.execute(stmt)).rowcount or 0
-            .values(
+        return counts["recovered"]
-                status="pending", lease_token=None, leased_at=None,
-                lease_expires_at=None, updated_at=now,
-            )
-        )
-        return res.rowcount or 0

backend/app/services/ml/gpu_triage.py

@@ -0,0 +1,156 @@
+"""GPU-failure triage (#125): classify errored jobs, PROBE the file, recover.
+
+An errored GPU job is a tombstone with a stored reason, but the reason alone is
+a suspicion, not a verdict — a timeout can hit a perfectly fine file, and
+"moov atom not found" can mean a truncated download OR a one-off transfer
+fault. So triage EVALUATES: it runs the real integrity probe (sha256 recompute
++ PIL/ffprobe — verify_integrity's own machinery) on each errored image ONCE
+and records both verdicts:
+
+  ImageRecord.integrity_status  <- file-level verdict (ok / corrupt / ...)
+  GpuJob.triage_status          <- 'defect'  (file is bad: recovery material,
+                                              excluded from /retry_errors)
+                                   'file_ok' (file passes: the failure was
+                                              operational, safe to retry)
+
+Recovery reuses established primitives: delete the defective copy + record
+(cleanup_service.delete_images — full cascade) and re-poll the image's
+subscription Source (the Layer-2 refetch pattern: gallery-dl re-fetches the
+now-absent file on the next source check). Images without a pollable Source
+report 'no_source' — manual remediation. Every classification is logged at
+WARNING so the operator notices in Logs / System Activity.
+"""
+import logging
+import time
+from datetime import UTC, datetime
+from pathlib import Path
+
+from sqlalchemy import select, update
+from sqlalchemy.orm import Session
+
+from ...models import GpuJob, ImageProvenance, ImageRecord, Source
+from ..cleanup_service import delete_images
+
+log = logging.getLogger(__name__)
+
+# Reason buckets for the triage overview (reporting only — the PROBE decides
+# 'defect', never the string). Ordered: first match wins.
+_REASON_BUCKETS = (
+    ("poisoned", ("poisoned:",)),
+    ("transient", ("gave up after repeated transient", "curator unreachable",
+                   "connection", "read timed out")),
+    ("timeout", ("timed out", "timeout")),
+    ("truncated_or_corrupt", ("moov atom", "invalid data", "end of file",
+                              "header missing", "error reading header",
+                              "truncated", "premature", "corrupt",
+                              "no frames sampled")),
+    ("decode", ("cannot identify", "decompression", "broken data stream",
+                "unrecognized data")),
+)
+
+
+def classify_reason(error: str | None) -> str:
+    """Bucket a stored job-error string for the overview table."""
+    text = (error or "").lower()
+    if not text:
+        return "other"
+    for bucket, needles in _REASON_BUCKETS:
+        if any(n in text for n in needles):
+            return bucket
+    return "other"
+
+
+def triage_errored_jobs(
+    session: Session, *, time_budget_seconds: float = 300.0,
+) -> dict:
+    """Probe every not-yet-triaged errored image and write both verdicts.
+
+    Time-boxed (sha256 of a large original over NFS can take tens of seconds)
+    and inherently resumable: rows are selected by `triage_status IS NULL`, so
+    the next sweep continues exactly where a budget cut stopped. Commits per
+    image so a mid-run crash keeps completed verdicts."""
+    image_ids = session.execute(
+        select(GpuJob.image_record_id)
+        .where(GpuJob.status == "error", GpuJob.triage_status.is_(None))
+        .group_by(GpuJob.image_record_id)
+        .order_by(GpuJob.image_record_id)
+    ).scalars().all()
+    counts = {"probed": 0, "defect": 0, "file_ok": 0, "partial": False}
+    if not image_ids:
+        return counts
+    # Lazy imports: the probe helper lives in the maintenance task module and
+    # the hasher in the importer — importing either at module load would pull
+    # celery into every service consumer.
+    from ...tasks.maintenance import _verify_one
+    from ..importer import _sha256_of
+
+    started = time.monotonic()
+    for image_id in image_ids:
+        if time.monotonic() - started > time_budget_seconds:
+            counts["partial"] = True
+            break
+        rec = session.get(ImageRecord, image_id)
+        if rec is None:      # record deleted since the job errored
+            continue
+        verdict = _verify_one(Path(rec.path), rec.sha256, rec.mime, _sha256_of)
+        # 'ok' means the failure was operational; anything else (corrupt /
+        # failed_verification = missing/unreadable) makes the file itself the
+        # problem — recovery material.
+        triage = "file_ok" if verdict == "ok" else "defect"
+        reason = session.execute(
+            select(GpuJob.error)
+            .where(GpuJob.image_record_id == image_id, GpuJob.status == "error")
+            .limit(1)
+        ).scalar_one_or_none()
+        rec.integrity_status = verdict
+        session.execute(
+            update(GpuJob)
+            .where(GpuJob.image_record_id == image_id, GpuJob.status == "error")
+            .values(triage_status=triage, updated_at=datetime.now(UTC))
+        )
+        session.commit()
+        counts["probed"] += 1
+        counts[triage] += 1
+        log.warning(
+            "gpu triage: image %s (%s) job error %r -> integrity probe %r -> %s",
+            image_id, rec.path, (reason or "")[:120], verdict, triage,
+        )
+    return counts
+
+
+def recover_defective_image(
+    session: Session, image_id: int, *, images_root: Path,
+) -> dict:
+    """Delete the defective copy + record and re-poll its subscription Source.
+
+    Mirrors the Layer-2 import refetch: with the bad file gone, the source's
+    next gallery-dl run re-fetches a fresh copy, which re-imports as a new
+    record and re-enters the GPU pipeline. The record delete cascades the
+    error tombstones with it. 'no_source' when no enabled, real-URL Source is
+    reachable via the image's provenance — manual remediation there."""
+    rec = session.get(ImageRecord, image_id)
+    if rec is None:
+        return {"status": "not_found"}
+    src_id = session.execute(
+        select(Source.id)
+        .join(ImageProvenance, ImageProvenance.source_id == Source.id)
+        .where(
+            ImageProvenance.image_record_id == image_id,
+            Source.enabled.is_(True),
+            ~Source.url.like("sidecar:%"),   # synthetic anchor — not pollable
+        )
+        .order_by(Source.id.asc())
+    ).scalars().first()
+    if src_id is None:
+        return {"status": "no_source"}
+    path = rec.path
+    summary = delete_images(session, image_ids=[image_id], images_root=images_root)
+    # Lazy import (services -> tasks would cycle at module load).
+    from ...tasks.download import download_source
+
+    download_source.delay(src_id)
+    log.warning(
+        "gpu triage recovery: deleted defective image %s (%s) and queued a "
+        "re-check of source %s to re-fetch it", image_id, path, src_id,
+    )
+    return {"status": "refetch_queued", "source_id": src_id, **summary}

backend/app/services/ml/heads.py

@@ -1,12 +1,13 @@
 """Production heads: train + score the per-concept classifiers (#114).
 
-The eval (#1130, tag_eval.py) proved the spine; this is its production form.
+The eval harness (#1130) proved the spine, then retired 2026-07-02 once the
+tagging system was accepted; this is the production form.
 - TRAIN (sync, ml worker — needs scikit-learn): for every general/character tag
   with enough labelled positives, fit a logistic-regression head on the FROZEN
   SigLIP embeddings (positives + negatives = rejections + sampled unlabeled),
   derive an honest suggest threshold + earned-auto-apply point from CROSS-
-  VALIDATED scores, and upsert a TagHead row. Reuses tag_eval's proven data
+  VALIDATED scores, and upsert a TagHead row. Uses the eval-proven data loaders
-  loaders + metric helpers so production heads match the eval's measured numbers.
+  + metric helpers (training_data.py) so heads match the measured numbers.
 - SCORE (async, API worker — numpy via pgvector, NO scikit-learn): score one
   image's embedding against all current heads → the suggestions the rail shows,
   REPLACING Camie predictions + per-tag centroids.
@@ -37,7 +38,7 @@ from ...models import (
     TagSuggestionRejection,
 )
 from ...models.tag import image_tag
-from .tag_eval import (
+from .training_data import (
     _auto_apply_point,
     _ids_with_tag,
     _l2norm,
@@ -308,25 +309,36 @@ async def score_image(
     import numpy as np
 
     img = await session.get(ImageRecord, image_id)
-    if img is None or img.siglip_embedding is None:
+    if img is None:
         return []
     settings = await _settings_async(session)
-    heads = await _current_heads(session, settings.embedder_model_version)
+    cur_version = settings.embedder_model_version
+    heads = await _current_heads(session, cur_version)
     if heads["W"] is None:
         return []
 
-    bag = [np.asarray(img.siglip_embedding, dtype=np.float32)]
+    # Only embeddings in the CURRENT model's space enter the bag. Mid model-swap
+    # (#1190), an image still carrying the OLD-version whole-image vector is
+    # skipped rather than scored by heads trained in a different space; a legacy
+    # NULL version is treated as current (those predate per-row stamping).
+    bag = []
+    if img.siglip_embedding is not None and img.siglip_model_version in (
+        cur_version, None,
+    ):
+        bag.append(np.asarray(img.siglip_embedding, dtype=np.float32))
     region_vecs = (
         await session.execute(
             select(ImageRegion.siglip_embedding)
             .where(ImageRegion.image_record_id == image_id)
             .where(ImageRegion.siglip_embedding.is_not(None))
-            .where(ImageRegion.embedding_version == settings.embedder_model_version)
+            .where(ImageRegion.embedding_version == cur_version)
         )
     ).all()
     for (vec,) in region_vecs:
         if vec is not None:
             bag.append(np.asarray(vec, dtype=np.float32))
+    if not bag:
+        return []
 
     X = np.vstack(bag)                                  # (B, D)
     norms = np.linalg.norm(X, axis=1, keepdims=True)

backend/app/services/ml/tag_eval.py

@@ -1,430 +0,0 @@
-"""Head-vs-centroid tagging eval (#1130, milestone #114 slice 1).
-
-Proves the "frozen embedding + small trained head (with negatives)" spine on the
-operator's OWN data, reusing the SigLIP embeddings already stored on
-image_record. For each concept tag it compares:
-  - CENTROID baseline (the old approach): cosine to the mean of positive vectors.
-  - HEAD (the new approach): logistic regression trained on positives + negatives.
-and reports cross-validated precision/recall/AP for both, a LEARNING CURVE
-(accuracy as the number of tagged positives grows), and example image ids to
-eyeball.
-
-numpy + scikit-learn are imported LAZILY inside run_eval so the API worker (base
-image, no ML stack) can still import start_tag_eval_run to enqueue the ml-queue
-task — the heavy compute only runs on the ml worker.
-"""
-
-from __future__ import annotations
-
-import logging
-from datetime import UTC, datetime
-from typing import Any
-
-from sqlalchemy import func, select
-from sqlalchemy.orm import Session
-
-from ...models import (
-    ImageRecord,
-    Tag,
-    TagEvalRun,
-    TagKind,
-    TagPositiveConfirmation,
-    TagSuggestionRejection,
-)
-from ...models.tag import image_tag
-
-log = logging.getLogger(__name__)
-
-# The operator's real concept list (mix of whole-ish + small/local cues). The
-# admin trigger can override; this is the default eval set.
-DEFAULT_CONCEPTS = [
-    "glasses", "cat", "dog", "horse", "goblin",
-    "cum", "lactation", "fellatio", "xray", "stomach bulge",
-]
-DEFAULT_CURVE_POINTS = [10, 30, 100, 300]
-DEFAULT_NEG_RATIO = 3       # negatives per positive (rejections + sampled unlabeled)
-DEFAULT_CV_FOLDS = 5
-MIN_POSITIVES = 8           # below this, a concept can't be evaluated meaningfully
-_UNLABELED_POOL = 4000      # cap on sampled unlabeled rows pulled per concept
-_EXAMPLES_K = 12
-
-
-def start_tag_eval_run(session: Session, params: dict[str, Any]) -> int:
-    """Create a TagEvalRun (status='running') and dispatch the ml-queue task.
-    Returns the new run id. Light guard: one running eval at a time."""
-    existing = session.execute(
-        select(TagEvalRun.id).where(TagEvalRun.status == "running")
-    ).scalar_one_or_none()
-    if existing is not None:
-        raise EvalAlreadyRunning(existing)
-    norm = _normalize_params(params)
-    run = TagEvalRun(params=norm, status="running", last_progress_at=datetime.now(UTC))
-    session.add(run)
-    session.flush()
-    run_id = run.id
-    # Same enqueue-by-import pattern api/suggestions.py uses for ml tasks; the
-    # commit happens in the API handler so row + dispatch are visible together.
-    from ...tasks.ml import tag_eval_run as _task
-    _task.delay(run_id)
-    return run_id
-
-
-class EvalAlreadyRunning(Exception):
-    """Raised by start_tag_eval_run when an eval is already in flight."""
-
-
-def _normalize_params(params: dict[str, Any] | None) -> dict[str, Any]:
-    params = params or {}
-    concepts = [str(c).strip() for c in (params.get("concepts") or []) if str(c).strip()]
-    try:
-        neg_ratio = max(1, int(params.get("neg_ratio", DEFAULT_NEG_RATIO)))
-    except (TypeError, ValueError):
-        neg_ratio = DEFAULT_NEG_RATIO
-    try:
-        cv_folds = max(2, int(params.get("cv_folds", DEFAULT_CV_FOLDS)))
-    except (TypeError, ValueError):
-        cv_folds = DEFAULT_CV_FOLDS
-    try:
-        auto_top_n = min(max(int(params.get("auto_top_n", 0) or 0), 0), 200)
-    except (TypeError, ValueError):
-        auto_top_n = 0
-    try:
-        precision_target = min(max(float(params.get("precision_target", 0.97)), 0.5), 0.999)
-    except (TypeError, ValueError):
-        precision_target = 0.97
-    # No explicit concepts and auto-discovery off → fall back to the hand list.
-    if not concepts and not auto_top_n:
-        concepts = list(DEFAULT_CONCEPTS)
-    curve = params.get("curve_points") or DEFAULT_CURVE_POINTS
-    curve = sorted({int(n) for n in curve if int(n) > 0})
-    return {
-        "concepts": concepts,
-        "neg_ratio": neg_ratio,
-        "cv_folds": cv_folds,
-        "auto_top_n": auto_top_n,
-        "precision_target": round(precision_target, 4),
-        "curve_points": curve,
-    }
-
-
-def _top_general_concepts(session: Session, n: int, min_count: int) -> list[str]:
-    """The n most-tagged general (concept) tags with >= min_count images — a fast
-    server-side way to broaden the eval beyond the hand-picked list (counts all
-    sources; source-aware filtering is a separate concern)."""
-    rows = session.execute(
-        select(Tag.name)
-        .join(image_tag, image_tag.c.tag_id == Tag.id)
-        .where(Tag.kind == TagKind.general)
-        .group_by(Tag.id)
-        .having(func.count(image_tag.c.image_record_id) >= min_count)
-        .order_by(func.count(image_tag.c.image_record_id).desc())
-        .limit(n)
-    ).all()
-    return [r[0] for r in rows]
-
-
-def _resolve_tag_id(session: Session, name: str) -> int | None:
-    """Case-insensitive tag-name match; if several share a name, take the one
-    applied to the most images (the one the operator actually uses)."""
-    rows = session.execute(
-        select(Tag.id, func.count(image_tag.c.image_record_id))
-        .outerjoin(image_tag, image_tag.c.tag_id == Tag.id)
-        .where(func.lower(Tag.name) == name.lower())
-        .group_by(Tag.id)
-        .order_by(func.count(image_tag.c.image_record_id).desc())
-    ).all()
-    return rows[0][0] if rows else None
-
-
-def _ids_with_tag(session: Session, tag_id: int) -> list[int]:
-    return [
-        r[0] for r in session.execute(
-            select(image_tag.c.image_record_id).where(image_tag.c.tag_id == tag_id)
-        ).all()
-    ]
-
-
-def _rejected_ids(session: Session, tag_id: int) -> list[int]:
-    return [
-        r[0] for r in session.execute(
-            select(TagSuggestionRejection.image_record_id)
-            .where(TagSuggestionRejection.tag_id == tag_id)
-        ).all()
-    ]
-
-
-def _confirmed_ids(session: Session, tag_id: int) -> set[int]:
-    """Positives the operator explicitly affirmed ('keep') — excluded from the
-    doubts list so confirmed-correct images don't resurface every run."""
-    return {
-        r[0] for r in session.execute(
-            select(TagPositiveConfirmation.image_record_id)
-            .where(TagPositiveConfirmation.tag_id == tag_id)
-        ).all()
-    }
-
-
-def _sample_unlabeled(session: Session, exclude: set[int], limit: int) -> list[int]:
-    """Random image ids (with an embedding) NOT carrying the tag. Concepts are
-    sparse, so an untagged image is almost always a true negative."""
-    stmt = (
-        select(ImageRecord.id)
-        .where(ImageRecord.siglip_embedding.is_not(None))
-        .order_by(func.random())
-        .limit(limit)
-    )
-    if exclude:
-        stmt = stmt.where(ImageRecord.id.not_in(exclude))
-    return [r[0] for r in session.execute(stmt).all()]
-
-
-def _load_embeddings(session: Session, ids: list[int]) -> dict[int, Any]:
-    import numpy as np
-
-    out: dict[int, Any] = {}
-    if not ids:
-        return out
-    # Chunk the IN list to stay well under psycopg's parameter ceiling.
-    for i in range(0, len(ids), 2000):
-        chunk = ids[i:i + 2000]
-        for rid, emb in session.execute(
-            select(ImageRecord.id, ImageRecord.siglip_embedding)
-            .where(ImageRecord.id.in_(chunk))
-            .where(ImageRecord.siglip_embedding.is_not(None))
-        ).all():
-            out[rid] = np.asarray(emb, dtype=np.float32)
-    return out
-
-
-def run_eval(session: Session, params: dict[str, Any]) -> dict[str, Any]:
-    """Compute the full report. Per-concept failures are captured, not fatal."""
-    import numpy as np
-
-    cfg = _normalize_params(params)
-    # Auto-discovery: union the explicit concepts with the top-N most-tagged
-    # general tags (server-side, fast) so the eval can broaden itself.
-    concepts = list(cfg["concepts"])
-    if cfg["auto_top_n"]:
-        seen = {c.lower() for c in concepts}
-        for name in _top_general_concepts(session, cfg["auto_top_n"], MIN_POSITIVES):
-            if name.lower() not in seen:
-                concepts.append(name)
-                seen.add(name.lower())
-    cfg["concepts"] = concepts
-    concepts_out = []
-    for name in cfg["concepts"]:
-        try:
-            concepts_out.append(_eval_concept(session, name, cfg, np))
-        except Exception as exc:  # one bad concept shouldn't kill the run
-            log.exception("tag-eval concept %r failed", name)
-            concepts_out.append({"name": name, "skipped": f"error: {exc}"})
-    return {
-        "generated_at": datetime.now(UTC).isoformat(),
-        "params": cfg,
-        "concepts": concepts_out,
-    }
-
-
-def _eval_concept(session: Session, name: str, cfg: dict, np) -> dict[str, Any]:
-    tag_id = _resolve_tag_id(session, name)
-    if tag_id is None:
-        return {"name": name, "skipped": "no such tag"}
-    pos_ids = _ids_with_tag(session, tag_id)
-    if len(pos_ids) < MIN_POSITIVES:
-        return {"name": name, "tag_id": tag_id, "n_pos": len(pos_ids),
-                "skipped": f"too few positives (<{MIN_POSITIVES})"}
-
-    neg_ratio = cfg["neg_ratio"]
-    pos_set = set(pos_ids)
-    rejected = [i for i in _rejected_ids(session, tag_id) if i not in pos_set]
-    want_neg = max(len(pos_ids) * neg_ratio, _EXAMPLES_K * 4)
-    sampled = _sample_unlabeled(session, pos_set | set(rejected),
-                                min(_UNLABELED_POOL, want_neg))
-    neg_ids = rejected + [i for i in sampled if i not in pos_set]
-
-    emb = _load_embeddings(session, pos_ids + neg_ids)
-    pos = [(i, emb[i]) for i in pos_ids if i in emb]
-    neg = [(i, emb[i]) for i in neg_ids if i in emb]
-    if len(pos) < MIN_POSITIVES or len(neg) < MIN_POSITIVES:
-        return {"name": name, "tag_id": tag_id, "n_pos": len(pos),
-                "n_neg": len(neg), "skipped": "too few embedded examples"}
-
-    ids = np.array([i for i, _ in pos] + [i for i, _ in neg])
-    X = np.vstack([v for _, v in pos] + [v for _, v in neg]).astype(np.float32)
-    y = np.array([1] * len(pos) + [0] * len(neg))
-    Xn = _l2norm(X, np)
-
-    head = _eval_head(Xn, y, cfg["cv_folds"], cfg["precision_target"], np)
-    centroid = _eval_centroid(Xn, y, cfg["cv_folds"], np)
-    curve = _learning_curve(Xn, y, cfg["curve_points"], neg_ratio, np)
-    confirmed = _confirmed_ids(session, tag_id)
-    examples = _examples(session, Xn, y, ids, np, set(rejected), confirmed)
-
-    return {
-        "name": name, "tag_id": tag_id,
-        "n_pos": len(pos), "n_neg": len(neg),
-        "n_rejected": len(rejected),
-        "head": head, "centroid": centroid,
-        "curve": curve, "examples": examples,
-    }
-
-
-def _l2norm(X, np):
-    n = np.linalg.norm(X, axis=1, keepdims=True)
-    n[n == 0] = 1.0
-    return X / n
-
-
-def _metrics_from_scores(y, scores, np) -> dict[str, float]:
-    from sklearn.metrics import average_precision_score, precision_recall_curve
-
-    ap = float(average_precision_score(y, scores))
-    prec, rec, thr = precision_recall_curve(y, scores)
-    f1 = (2 * prec * rec) / np.clip(prec + rec, 1e-9, None)
-    best = int(np.argmax(f1))
-    # thr has len = len(prec)-1; map best index safely.
-    t = float(thr[min(best, len(thr) - 1)]) if len(thr) else 0.5
-    return {
-        "ap": round(ap, 4),
-        "precision": round(float(prec[best]), 4),
-        "recall": round(float(rec[best]), 4),
-        "f1": round(float(f1[best]), 4),
-        "threshold": round(t, 4),
-    }
-
-
-def _safe_folds(y, folds, np) -> int:
-    minority = int(min(np.bincount(y)))
-    return max(2, min(folds, minority))
-
-
-def _eval_head(Xn, y, folds, target, np) -> dict[str, float]:
-    from sklearn.linear_model import LogisticRegression
-    from sklearn.model_selection import StratifiedKFold, cross_val_predict
-
-    clf = LogisticRegression(max_iter=1000, class_weight="balanced")
-    cv = StratifiedKFold(n_splits=_safe_folds(y, folds, np), shuffle=True,
-                         random_state=0)
-    probs = cross_val_predict(clf, Xn, y, cv=cv, method="predict_proba")[:, 1]
-    m = _metrics_from_scores(y, probs, np)
-    m["auto_apply"] = _auto_apply_point(y, probs, target, np)
-    return m
-
-
-def _auto_apply_point(y, scores, target, np) -> dict | None:
-    """The auto-apply operating point: the threshold that yields the MOST recall
-    while holding precision >= target. This answers 'could this concept fire
-    without a human, and how much would it catch?' Returns None if no threshold
-    reaches the precision target (concept not auto-apply-ready)."""
-    from sklearn.metrics import precision_recall_curve
-
-    prec, rec, thr = precision_recall_curve(y, scores)
-    best = None  # (threshold, precision, recall) maximizing recall s.t. prec>=target
-    for i in range(len(thr)):  # thr[i] corresponds to prec[i], rec[i]
-        if prec[i] >= target and (best is None or rec[i] > best[2]):
-            best = (float(thr[i]), float(prec[i]), float(rec[i]))
-    if best is None:
-        return None
-    return {
-        "target": round(float(target), 4),
-        "threshold": round(best[0], 4),
-        "precision": round(best[1], 4),
-        "recall": round(best[2], 4),
-    }
-
-
-def _eval_centroid(Xn, y, folds, np) -> dict[str, float]:
-    """Cross-validated cosine-to-positive-mean — the OLD method's quality."""
-    from sklearn.model_selection import StratifiedKFold
-
-    cv = StratifiedKFold(n_splits=_safe_folds(y, folds, np), shuffle=True,
-                         random_state=0)
-    scores = np.zeros(len(y), dtype=np.float32)
-    for train, test in cv.split(Xn, y):
-        c = Xn[train][y[train] == 1].mean(axis=0)
-        cn = c / (np.linalg.norm(c) or 1.0)
-        scores[test] = Xn[test] @ cn
-    return _metrics_from_scores(y, scores, np)
-
-
-def _learning_curve(Xn, y, points, neg_ratio, np) -> list[dict[str, float]]:
-    """Hold out a fixed test split; train the head on a growing number of
-    positives and watch AP/F1 climb — answers 'does tagging more sharpen it?'"""
-    from sklearn.linear_model import LogisticRegression
-    from sklearn.model_selection import train_test_split
-
-    rng = np.random.default_rng(0)
-    idx = np.arange(len(y))
-    try:
-        tr, te = train_test_split(idx, test_size=0.3, stratify=y, random_state=0)
-    except ValueError:
-        return []
-    tr_pos = tr[y[tr] == 1]
-    tr_neg = tr[y[tr] == 0]
-    out = []
-    for n in points:
-        if n > len(tr_pos):
-            break
-        sp = rng.choice(tr_pos, size=n, replace=False)
-        nn = min(len(tr_neg), n * neg_ratio)
-        sn = rng.choice(tr_neg, size=nn, replace=False)
-        sub = np.concatenate([sp, sn])
-        clf = LogisticRegression(max_iter=1000, class_weight="balanced")
-        clf.fit(Xn[sub], y[sub])
-        prob = clf.predict_proba(Xn[te])[:, 1]
-        m = _metrics_from_scores(y[te], prob, np)
-        out.append({"n_pos": int(n), "ap": m["ap"], "f1": m["f1"]})
-    return out
-
-
-def _examples(session, Xn, y, ids, np, rejected_set, confirmed_set) -> dict[str, list[dict]]:
-    """Train on all data, then surface: top-scoring negatives the operator has
-    NOT already rejected (= fresh suggestions) and lowest-scoring POSITIVES the
-    operator has NOT already confirmed (= unreviewed doubts). Excluding rejected
-    ids stops an adjudicated near-miss from resurfacing in 'would suggest';
-    excluding confirmed ids stops a 'kept' correct positive from resurfacing in
-    'head doubts' every run. Resolves thumbnail urls for a self-contained report."""
-    from sklearn.linear_model import LogisticRegression
-
-    clf = LogisticRegression(max_iter=1000, class_weight="balanced")
-    clf.fit(Xn, y)
-    s = clf.predict_proba(Xn)[:, 1]
-    neg_idx = np.where(y == 0)[0]
-    pos_idx = np.where(y == 1)[0]
-    top_neg = []
-    for i in neg_idx[np.argsort(s[neg_idx])[::-1]]:  # high score → low
-        rid = int(ids[i])
-        if rid in rejected_set:
-            continue  # already told the head 'no' — don't re-suggest it
-        top_neg.append(rid)
-        if len(top_neg) >= _EXAMPLES_K:
-            break
-    low_pos = []
-    for i in pos_idx[np.argsort(s[pos_idx])]:  # low score → high
-        rid = int(ids[i])
-        if rid in confirmed_set:
-            continue  # already kept/confirmed — don't re-doubt it
-        low_pos.append(rid)
-        if len(low_pos) >= _EXAMPLES_K:
-            break
-    thumbs = _resolve_thumbs(session, top_neg + low_pos)
-    return {
-        "head_would_suggest": [thumbs[i] for i in top_neg if i in thumbs],
-        "head_doubts_positive": [thumbs[i] for i in low_pos if i in thumbs],
-    }
-
-
-def _resolve_thumbs(session, ids: list[int]) -> dict[int, dict]:
-    from ..gallery_service import thumbnail_url
-
-    out: dict[int, dict] = {}
-    if not ids:
-        return out
-    for rid, tp, sha, mime in session.execute(
-        select(
-            ImageRecord.id, ImageRecord.thumbnail_path,
-            ImageRecord.sha256, ImageRecord.mime,
-        ).where(ImageRecord.id.in_(ids))
-    ).all():
-        out[rid] = {"id": rid, "thumbnail_url": thumbnail_url(tp, sha, mime)}
-    return out

backend/app/services/ml/tagger.py

@@ -1,210 +0,0 @@
-"""Camie-tagger-v2 ONNX wrapper (CPU).
-
-Single-image at a time. Loaded lazily inside the ml-worker process; NOT
-thread-safe — the ml queue worker runs --concurrency=1 per process (scale ML by
-running multiple worker replicas, not threads).
-
-v2 layout reference: HuggingFace Camais03/camie-tagger-v2 root has
-camie-tagger-v2.onnx (789 MB) + camie-tagger-v2-metadata.json (7.77 MB)
-+ config.json. Tags ship as nested JSON, not CSV. Preprocessing and
-output handling follow the published onnx_inference.py reference:
-ImageNet normalize, NCHW layout, sigmoid on refined logits (output[1]).
-"""
-
-import json
-import os
-from dataclasses import dataclass
-from pathlib import Path
-
-import numpy as np
-from PIL import Image, ImageFile
-
-# Cap inference threads (see Tagger.load) so each ml-worker replica is a bounded
-# core consumer on a shared node — keep N_replicas × this within the cores
-# allotted to ML so replicas don't oversubscribe the box / starve the DB.
-_INTRA_OP_THREADS = 4
-
-# onnxruntime lives in requirements-ml.txt only — it is NOT installed in the
-# lean web image or in CI. Imported lazily inside Tagger.load() so this module
-# imports fine without it (the suggestion service imports SURFACED_CATEGORIES
-# from here in the web container, and CI collects the pure-logic tests).
-
-# Tolerate minutely-truncated source images (same rationale as IR's wd14.py:
-# a few missing bytes at the JPEG EOI shouldn't block tagging the whole image).
-ImageFile.LOAD_TRUNCATED_IMAGES = True
-
-MODEL_NAME = os.environ.get("CAMIE_MODEL_NAME", "camie-tagger-v2")
-_MODEL_DIR = Path(os.environ.get("ML_MODEL_DIR", "/models")) / "camie"
-_MODEL_FILE = f"{MODEL_NAME}.onnx"
-_METADATA_FILE = f"{MODEL_NAME}-metadata.json"
-
-# Ingest floor below which predictions aren't stored (keeps the JSON compact).
-# DEFAULT/fallback only — the live value is DB-backed
-# (ml_settings.tagger_store_floor) and passed into infer() per call by the ml
-# task. 0.70: the suggestion path already filters there and the centroid path
-# covers lower-confidence preferred tags, so the sub-0.70 tail is redundant
-# (it had bloated image_record's TOAST to ~100 GB; plan-task #764).
-DEFAULT_STORE_FLOOR = 0.70
-
-# The categories FC-2b surfaces in the UI. Others (meta/rating/year) are
-# still stored but the suggestion service filters them out.
-# 'artist' retired in FC-2d-vii-c — artist identity is acquisition-derived
-# (image_record.artist_id), never ML-inferred. 'copyright' retired
-# 2026-06-01 — operator doesn't use the copyright tag-kind; fandom is
-# this app's franchise/series concept (per TagsView.vue's doc comment).
-# Raw predictions for both categories still get stored at STORE_FLOOR but
-# don't surface in suggestions.
-SURFACED_CATEGORIES = {"character", "general"}
-
-# ImageNet preprocessing constants (per Camie v2 onnx_inference.py).
-_IMAGENET_MEAN = np.array([0.485, 0.456, 0.406], dtype=np.float32)
-_IMAGENET_STD = np.array([0.229, 0.224, 0.225], dtype=np.float32)
-# Square-pad color ≈ ImageNet mean × 255 (matches reference inference).
-_PAD_COLOR = (124, 116, 104)
-
-
-@dataclass(frozen=True)
-class TagPrediction:
-    name: str
-    category: str
-    confidence: float
-
-
-class Tagger:
-    def __init__(self, model_dir: Path | None = None):
-        self._model_dir = model_dir or _MODEL_DIR
-        self._session = None  # onnxruntime.InferenceSession once load()ed
-        self._tag_names: list[str] | None = None
-        self._tag_categories: list[str] | None = None
-        self._input_name: str | None = None
-        self._input_size: int = 512
-
-    def load(self) -> None:
-        if self._session is not None:
-            return
-        model_path = self._model_dir / _MODEL_FILE
-        meta_path = self._model_dir / _METADATA_FILE
-        if not model_path.is_file():
-            raise RuntimeError(
-                f"Camie {_MODEL_FILE} missing at {model_path}. "
-                f"Populate /models via the ml-worker downloader."
-            )
-        if not meta_path.is_file():
-            raise RuntimeError(
-                f"Camie {_METADATA_FILE} missing at {meta_path}. "
-                f"Populate /models via the ml-worker downloader."
-            )
-
-        with open(meta_path) as f:
-            metadata = json.load(f)
-
-        # Per Camie v2 onnx_inference.py: idx_to_tag is keyed by str(idx);
-        # tag_to_category maps tag_name -> category. Project to two parallel
-        # lists indexed by output position for O(1) lookup in the hot path.
-        ds = metadata["dataset_info"]
-        idx_to_tag = ds["tag_mapping"]["idx_to_tag"]
-        tag_to_category = ds["tag_mapping"]["tag_to_category"]
-        total = ds["total_tags"]
-        names: list[str] = []
-        cats: list[str] = []
-        for i in range(total):
-            name = idx_to_tag.get(str(i), f"unknown-{i}")
-            names.append(name)
-            cats.append(tag_to_category.get(name, "general"))
-
-        # Input size from metadata; fall back to 512 (the v2 default).
-        self._input_size = int(
-            metadata.get("model_info", {}).get("img_size", 512)
-        )
-
-        # Lazy import — kept after the file-existence checks so the
-        # missing-model RuntimeError still fires first in environments
-        # without onnxruntime (CI / lean web image).
-        import onnxruntime as ort
-
-        # Cap the intra-op thread pool. ONNX Runtime otherwise sizes it to ALL
-        # host cores, so on a shared node each ml-worker replica would grab every
-        # core and oversubscribe (and starve the co-located DB/web). Bounding it
-        # makes each replica a predictable core consumer — run N replicas where
-        # N × _INTRA_OP_THREADS stays within the cores you allot to ML.
-        opts = ort.SessionOptions()
-        opts.intra_op_num_threads = _INTRA_OP_THREADS
-        session = ort.InferenceSession(
-            str(model_path), sess_options=opts, providers=["CPUExecutionProvider"],
-        )
-        self._input_name = session.get_inputs()[0].name
-        # Assign sentinels last so a partial load isn't observable.
-        self._tag_names = names
-        self._tag_categories = cats
-        self._session = session
-
-    def _preprocess(self, image_path: Path) -> np.ndarray:
-        img = Image.open(image_path)
-        # Composite RGBA onto neutral so transparency doesn't bias the model.
-        if img.mode == "RGBA":
-            bg = Image.new("RGBA", img.size, (255, 255, 255, 255))
-            bg.paste(img, mask=img.split()[3])
-            img = bg.convert("RGB")
-        elif img.mode != "RGB":
-            img = img.convert("RGB")
-
-        # Pad to square with ImageNet-mean color, then bicubic resize.
-        w, h = img.size
-        side = max(w, h)
-        square = Image.new("RGB", (side, side), _PAD_COLOR)
-        square.paste(img, ((side - w) // 2, (side - h) // 2))
-        square = square.resize(
-            (self._input_size, self._input_size), Image.BICUBIC
-        )
-
-        arr = np.array(square, dtype=np.float32) / 255.0   # HWC, [0,1]
-        arr = (arr - _IMAGENET_MEAN) / _IMAGENET_STD        # ImageNet normalize
-        arr = arr.transpose(2, 0, 1)                        # HWC -> CHW
-        return arr[np.newaxis, :, :, :]                     # NCHW
-
-    def infer(
-        self, image_path: Path, *, store_floor: float = DEFAULT_STORE_FLOOR,
-    ) -> dict[str, TagPrediction]:
-        """Run Camie v2 on one image. Returns {name: TagPrediction} with
-        confidence >= store_floor (across all categories — the suggestion
-        service does category filtering later). store_floor is the DB-backed
-        ml_settings.tagger_store_floor, passed in by the ml task.
-
-        v2 emits multiple outputs; we use the refined predictions
-        (output[1] per onnx_inference.py). Sigmoid is applied to raw
-        logits to produce [0,1] confidence scores.
-        """
-        self.load()
-        x = self._preprocess(image_path)
-        outputs = self._session.run(None, {self._input_name: x})
-        # Refined predictions if present (v2 emits initial + refined),
-        # fall back to initial for single-output forks.
-        logits = outputs[1] if len(outputs) > 1 else outputs[0]
-        # Squeeze batch dim, apply sigmoid.
-        probs = 1.0 / (1.0 + np.exp(-logits[0]))
-        results: dict[str, TagPrediction] = {}
-        names = self._tag_names
-        cats = self._tag_categories
-        for idx, score in enumerate(probs):
-            conf = float(score)
-            if conf < store_floor:
-                continue
-            if idx >= len(names):
-                # Output longer than metadata declared — shouldn't happen but
-                # don't crash the import pipeline if v2 metadata desynchronizes.
-                continue
-            results[names[idx]] = TagPrediction(
-                name=names[idx], category=cats[idx], confidence=conf
-            )
-        return results
-
-
-_default_tagger: Tagger | None = None
-
-
-def get_tagger() -> Tagger:
-    """Process-level singleton so the ONNX session loads once per worker."""
-    global _default_tagger
-    if _default_tagger is None:
-        _default_tagger = Tagger()
-    return _default_tagger

backend/app/services/ml/training_data.py

@@ -0,0 +1,121 @@
+"""Shared data-selection + validated-metric helpers for the heads trainer.
+
+Born in the head-vs-centroid eval harness (#1130, tag_eval.py) that proved the
+"frozen embedding + small trained head (with negatives)" spine; the harness was
+retired 2026-07-02 (operator: the tagging system is proven, the eval isn't
+needed) and these survivors moved here — they ARE the heads' production data
+pipeline (heads.py trains and scores with them nightly).
+
+numpy/scikit-learn are imported lazily inside the functions that need them so
+the API worker (base image, no ML stack) can import this module.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from sqlalchemy import func, select
+from sqlalchemy.orm import Session
+
+from ...models import ImageRecord, TagSuggestionRejection
+from ...models.tag import image_tag
+
+
+def _ids_with_tag(session: Session, tag_id: int) -> list[int]:
+    return [
+        r[0] for r in session.execute(
+            select(image_tag.c.image_record_id).where(image_tag.c.tag_id == tag_id)
+        ).all()
+    ]
+
+
+def _rejected_ids(session: Session, tag_id: int) -> list[int]:
+    return [
+        r[0] for r in session.execute(
+            select(TagSuggestionRejection.image_record_id)
+            .where(TagSuggestionRejection.tag_id == tag_id)
+        ).all()
+    ]
+
+
+def _sample_unlabeled(session: Session, exclude: set[int], limit: int) -> list[int]:
+    """Random image ids (with an embedding) NOT carrying the tag. Concepts are
+    sparse, so an untagged image is almost always a true negative."""
+    stmt = (
+        select(ImageRecord.id)
+        .where(ImageRecord.siglip_embedding.is_not(None))
+        .order_by(func.random())
+        .limit(limit)
+    )
+    if exclude:
+        stmt = stmt.where(ImageRecord.id.not_in(exclude))
+    return [r[0] for r in session.execute(stmt).all()]
+
+
+def _load_embeddings(session: Session, ids: list[int]) -> dict[int, Any]:
+    import numpy as np
+
+    out: dict[int, Any] = {}
+    if not ids:
+        return out
+    # Chunk the IN list to stay well under psycopg's parameter ceiling.
+    for i in range(0, len(ids), 2000):
+        chunk = ids[i:i + 2000]
+        for rid, emb in session.execute(
+            select(ImageRecord.id, ImageRecord.siglip_embedding)
+            .where(ImageRecord.id.in_(chunk))
+            .where(ImageRecord.siglip_embedding.is_not(None))
+        ).all():
+            out[rid] = np.asarray(emb, dtype=np.float32)
+    return out
+
+
+def _l2norm(X, np):
+    n = np.linalg.norm(X, axis=1, keepdims=True)
+    n[n == 0] = 1.0
+    return X / n
+
+
+def _metrics_from_scores(y, scores, np) -> dict[str, float]:
+    from sklearn.metrics import average_precision_score, precision_recall_curve
+
+    ap = float(average_precision_score(y, scores))
+    prec, rec, thr = precision_recall_curve(y, scores)
+    f1 = (2 * prec * rec) / np.clip(prec + rec, 1e-9, None)
+    best = int(np.argmax(f1))
+    # thr has len = len(prec)-1; map best index safely.
+    t = float(thr[min(best, len(thr) - 1)]) if len(thr) else 0.5
+    return {
+        "ap": round(ap, 4),
+        "precision": round(float(prec[best]), 4),
+        "recall": round(float(rec[best]), 4),
+        "f1": round(float(f1[best]), 4),
+        "threshold": round(t, 4),
+    }
+
+
+def _safe_folds(y, folds, np) -> int:
+    minority = int(min(np.bincount(y)))
+    return max(2, min(folds, minority))
+
+
+def _auto_apply_point(y, scores, target, np) -> dict | None:
+    """The auto-apply operating point: the threshold that yields the MOST recall
+    while holding precision >= target. This answers 'could this concept fire
+    without a human, and how much would it catch?' Returns None if no threshold
+    reaches the precision target (concept not auto-apply-ready)."""
+    from sklearn.metrics import precision_recall_curve
+
+    prec, rec, thr = precision_recall_curve(y, scores)
+    best = None  # (threshold, precision, recall) maximizing recall s.t. prec>=target
+    for i in range(len(thr)):  # thr[i] corresponds to prec[i], rec[i]
+        if prec[i] >= target and (best is None or rec[i] > best[2]):
+            best = (float(thr[i]), float(prec[i]), float(rec[i]))
+    if best is None:
+        return None
+    return {
+        "target": round(float(target), 4),
+        "threshold": round(best[0], 4),
+        "precision": round(best[1], 4),
+        "recall": round(best[2], 4),
+    }

backend/app/services/tag_service.py

@@ -10,8 +10,6 @@ from sqlalchemy.dialects.postgresql import insert as pg_insert
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from ..models import HeadMetric, Tag, TagHead, TagKind, image_tag
-from ..models.tag_allowlist import TagAllowlist
-from ..models.tag_reference_embedding import TagReferenceEmbedding
 from .db_helpers import get_or_create
 from .tag_query import fandom_join_alias, tag_columns
 
@@ -304,35 +302,22 @@ class TagService:
 
     async def _keep_as_alias(self, tag_id: int) -> bool:
         """A merged-away tag's old name must survive as an alias iff the ML
-        pipeline has ever applied it OR could re-emit it (allowlisted / has
+        pipeline has ever applied it (manual accept or head auto-apply) — so a
-        a centroid) — otherwise the proactive apply_allowlist_tags worker
+        re-application or an alias remap resolves the canonical name. Purely-
-        would silently regenerate it. Purely-manual, ML-unknown tags are
+        manual, ML-unknown tags are deleted outright (no DB bloat)."""
-        deleted outright (no DB bloat)."""
         is_machine = await self.session.scalar(
             select(
                 exists().where(
                     and_(
                         image_tag.c.tag_id == tag_id,
                         image_tag.c.source.in_(
-                            ("ml_auto", "ml_accepted", "auto")
+                            ("ml_auto", "ml_accepted", "head_auto", "auto")
                         ),
                     )
                 )
             )
         )
-        if is_machine:
+        return bool(is_machine)
-            return True
-        allowlisted = await self.session.scalar(
-            select(exists().where(TagAllowlist.tag_id == tag_id))
-        )
-        if allowlisted:
-            return True
-        has_centroid = await self.session.scalar(
-            select(
-                exists().where(TagReferenceEmbedding.tag_id == tag_id)
-            )
-        )
-        return bool(has_centroid)
 
     async def rename(self, tag_id: int, new_name: str) -> Tag:
         """Rename a tag. Raises TagMergeConflict if the new name collides
@@ -572,8 +557,6 @@ class TagService:
 
         merged_count = await self._repoint_image_tags(source_id, target_id)
         await self._repoint_rejections(source_id, target_id)
-        await self._repoint_allowlist(source_id, target_id)
-        await self._repoint_embedding(source_id)
         await self._repoint_aliases(source_id, target_id)
         await self._repoint_fandom_children(
             source_id, target_id, source_kind
@@ -639,30 +622,6 @@ class TagService:
             .values(tag_id=tgt)
         )
 
-    async def _repoint_allowlist(self, src: int, tgt: int) -> None:
-        tgt_has = await self.session.scalar(
-            select(exists().where(TagAllowlist.tag_id == tgt))
-        )
-        if tgt_has:
-            await self.session.execute(
-                text("DELETE FROM tag_allowlist WHERE tag_id = :src"),
-                {"src": src},
-            )
-        else:
-            await self.session.execute(
-                update(TagAllowlist)
-                .where(TagAllowlist.tag_id == src)
-                .values(tag_id=tgt)
-            )
-
-    async def _repoint_embedding(self, src: int) -> None:
-        await self.session.execute(
-            text(
-                "DELETE FROM tag_reference_embedding WHERE tag_id = :src"
-            ),
-            {"src": src},
-        )
-
     async def _repoint_aliases(self, src: int, tgt: int) -> None:
         from ..models.tag_alias import TagAlias
 

backend/app/tasks/external.py

@@ -216,11 +216,13 @@ def fetch_external_link(self, link_id: int, _serialize_waits: int = 0) -> dict:
         # Thumbnails + ML for any newly-attached images (mirrors the download
         # path). Lazy import to dodge a task-module import cycle.
         if image_ids:
-            from .ml import tag_and_embed
+            from .ml import cpu_embed_enabled, embed_image
             from .thumbnail import generate_thumbnail
+            do_embed = cpu_embed_enabled()
             for img_id in image_ids:
                 generate_thumbnail.delay(img_id)
-                tag_and_embed.delay(img_id)
+                if do_embed:
+                    embed_image.delay(img_id)
         return {"link_id": link_id, "files": len(result.files), "images": len(image_ids)}
     except Exception as exc:  # never leave a link stuck in 'downloading'
         log.exception("external fetch task failed for link %s", link_id)

backend/app/tasks/gpu_queue.py

@@ -0,0 +1,171 @@
+"""GPU-job queue coordination: backfill enqueues, orphan recovery, reprocess.
+
+These are pure-DB sweeps (INSERT…SELECT / UPDATE) — no torch, no sklearn —
+that keep the desktop GPU agent's work queue fed and self-healing. They lived
+in tasks/ml.py (routed to the 'ml' queue) purely by colocation, which made the
+ml-worker container a hard dependency of the GPU pipeline; under B3 the
+ml-worker is OPTIONAL (its only processing role is the CPU embed fallback), so
+these moved here and route to the 'maintenance' quick lane with the other
+recovery sweeps. A stack with no ml-worker keeps a fully-working GPU pipeline.
+"""
+
+import logging
+
+from sqlalchemy import select
+
+from ..celery_app import celery
+from ._sync_engine import sync_session_factory as _sync_session_factory
+
+log = logging.getLogger(__name__)
+
+
+@celery.task(name="backend.app.tasks.gpu_queue.enqueue_gpu_backfill")
+def enqueue_gpu_backfill(task_name: str) -> int:
+    """Enqueue a gpu_job for every image that still needs `task_name` (one
+    INSERT…SELECT, so it scales to a full library). The desktop agent drains the
+    queue over HTTP. Returns the number enqueued.
+
+    Completion is judged PER PIPELINE, never across them (B3, operator
+    2026-07-02): 'ccip' by prior gpu_job rows, 'siglip' by concept regions at
+    the current model version, and only 'embed' by image_record's whole-image
+    embedding — the one artifact the CPU fallback also produces. A CPU embed
+    therefore never closes crop/detect work for the agent.
+
+    An ERRORED job is a tombstone for its (image, task): no variant re-enqueues
+    it. Retry is deliberate-only (/retry_errors), which also means an errored
+    back-catalogue needs one "Retry errored jobs" press after a model swap.
+    Before the tombstone rule, this loop re-minted a fresh doomed job for every
+    permanently-bad file each run — ~24 duplicate error rows/day per file (the
+    2026-07-02 "unprocessable" flood)."""
+    from sqlalchemy import exists, insert, literal, or_
+    from sqlalchemy import select as sa_select
+
+    from ..models import GpuJob, ImageRecord, ImageRegion, MLSettings
+    from ..services.ml.gpu_jobs import error_dedupe_statements
+
+    SessionLocal = _sync_session_factory()
+    with SessionLocal() as session:
+        # Prune stale tombstones first (loop-era duplicates + rows made moot by
+        # a later success), so 'error' reads as one row per distinct failing
+        # file and the skip-guards below see a clean picture.
+        pruned = sum(
+            session.execute(s).rowcount or 0 for s in error_dedupe_statements()
+        )
+        if pruned:
+            log.info("gpu backfill: pruned %d stale/duplicate error rows", pruned)
+        cur_version = session.execute(
+            select(MLSettings.embedder_model_version).where(MLSettings.id == 1)
+        ).scalar_one()
+        if task_name == "embed":
+            # Whole-image GPU re-embed (#1190): images with no embedding, or one
+            # stamped under a DIFFERENT model version (an operator model swap).
+            stale = or_(
+                ImageRecord.siglip_embedding.is_(None),
+                ImageRecord.siglip_model_version.is_(None),
+                ImageRecord.siglip_model_version != cur_version,
+            )
+            # 'error' blocks too — tombstone rule, see docstring.
+            blocked = exists().where(
+                GpuJob.image_record_id == ImageRecord.id,
+                GpuJob.task == "embed",
+                GpuJob.status.in_(["pending", "leased", "error"]),
+            )
+            sel = sa_select(
+                ImageRecord.id, literal("embed"), literal("pending")
+            ).where(stale).where(~blocked)
+        elif task_name == "siglip":
+            # Concept-crop re-embed: enqueue when there's no concept region AT THE
+            # CURRENT model version — so a model swap re-triggers crops too, not
+            # only the never-embedded back-catalogue.
+            has_current_concept = exists().where(
+                ImageRegion.image_record_id == ImageRecord.id,
+                ImageRegion.kind == "concept",
+                ImageRegion.embedding_version == cur_version,
+            )
+            # 'error' blocks too — tombstone rule, see docstring.
+            blocked = exists().where(
+                GpuJob.image_record_id == ImageRecord.id,
+                GpuJob.task == "siglip",
+                GpuJob.status.in_(["pending", "leased", "error"]),
+            )
+            sel = sa_select(
+                ImageRecord.id, literal("siglip"), literal("pending")
+            ).where(~has_current_concept).where(~blocked)
+        else:
+            # ANY prior row blocks — including 'error' (tombstone rule, see
+            # docstring): pre-fix this branch ran HOURLY and was the loop.
+            already = exists().where(
+                GpuJob.image_record_id == ImageRecord.id,
+                GpuJob.task == task_name,
+                GpuJob.status.in_(["pending", "leased", "done", "error"]),
+            )
+            sel = sa_select(
+                ImageRecord.id, literal(task_name), literal("pending")
+            ).where(~already)
+        # RETURNING + count: result.rowcount is unreliable for INSERT…SELECT.
+        rows = session.execute(
+            insert(GpuJob)
+            .from_select(["image_record_id", "task", "status"], sel)
+            .returning(GpuJob.id)
+        ).fetchall()
+        session.commit()
+        return len(rows)
+
+
+@celery.task(name="backend.app.tasks.gpu_queue.recover_orphaned_gpu_jobs")
+def recover_orphaned_gpu_jobs() -> int:
+    """Reset expired GPU-job leases back to pending — recovers work orphaned by an
+    agent that died mid-job (no graceful release) — and convert poison-loopers
+    (release/expiry cycles that never reach fail()'s attempt cap) to 'error'.
+    Statements are shared with GpuJobService.recover_orphaned so the sweep and
+    the service can't drift. Short beat cadence so orphans get picked back up
+    quickly + the queue counts read honestly. Returns the number recovered."""
+    from datetime import UTC, datetime
+
+    from ..services.ml.gpu_jobs import recover_statements
+
+    SessionLocal = _sync_session_factory()
+    with SessionLocal() as session:
+        counts = {
+            name: session.execute(stmt).rowcount or 0
+            for name, stmt in recover_statements(datetime.now(UTC)).items()
+        }
+        session.commit()
+        if counts["poison_expired"] or counts["poison_pending"]:
+            log.warning(
+                "gpu jobs poisoned -> error: %d crash-loop (expired lease), "
+                "%d never-complete (pending)",
+                counts["poison_expired"], counts["poison_pending"],
+            )
+        return counts["recovered"]
+
+
+@celery.task(name="backend.app.tasks.gpu_queue.reprocess_gpu_jobs")
+def reprocess_gpu_jobs(task_name: str = "ccip") -> int:
+    """Reset every done/error job of `task_name` back to pending so the agent
+    re-runs the WHOLE library under the CURRENT pipeline — e.g. after adding crop
+    detectors (#1202), re-cropping existing images. Heavy + operator-triggered;
+    the back-catalogue won't otherwise re-process (the backfills skip images that
+    already have current-version regions). Returns the number reset."""
+    from datetime import UTC, datetime
+
+    from sqlalchemy import update
+
+    from ..models import GpuJob
+
+    SessionLocal = _sync_session_factory()
+    with SessionLocal() as session:
+        now = datetime.now(UTC)
+        res = session.execute(
+            update(GpuJob)
+            .where(
+                GpuJob.task == task_name,
+                GpuJob.status.in_(["done", "error"]),
+            )
+            .values(
+                status="pending", attempts=0, lease_token=None, leased_at=None,
+                lease_expires_at=None, updated_at=now,
+            )
+        )
+        session.commit()
+        return res.rowcount or 0

backend/app/tasks/import_file.py

@@ -228,15 +228,17 @@ def _do_import(session, task, import_task_id: int) -> dict:
     # Enqueue thumbnail + ML for newly imported AND superseded images
     # (a superseded row has cleared ML + no thumbnail).
     if result.status in ("imported", "superseded"):
-        from .ml import tag_and_embed
+        from .ml import cpu_embed_enabled, embed_image
         from .thumbnail import generate_thumbnail
 
+        do_embed = cpu_embed_enabled()
         ids = list(result.member_image_ids)
         if result.image_id is not None and result.image_id not in ids:
             ids.append(result.image_id)
         for img_id in ids:
             generate_thumbnail.delay(img_id)
-            tag_and_embed.delay(img_id)
+            if do_embed:
+                embed_image.delay(img_id)
 
     # If this was the last task in the batch, mark the batch complete.
     remaining = session.execute(

backend/app/tasks/maintenance.py

@@ -21,7 +21,6 @@ from ..models import (
     ImportTask,
     LibraryAuditRun,
     Source,
-    TagEvalRun,
     TaskRun,
 )
 from ..utils.phash import compute_phash
@@ -96,9 +95,6 @@ BACKUP_DB_STALL_THRESHOLD_MINUTES = 40
 # Library audit: scan_library_for_rule has time_limit=7500s (2h5m).
 # 2h15m gives a 10-min buffer.
 LIBRARY_AUDIT_STALL_THRESHOLD_MINUTES = 135
-# tag-eval (#1130) has a 30-min soft limit; flag a run with no progress past 40.
-TAG_EVAL_STALL_THRESHOLD_MINUTES = 40
-TAG_EVAL_KEEP_RUNS = 20
 # head training (#114) has a 60-min soft limit; flag no-progress past 75.
 HEAD_TRAINING_STALL_THRESHOLD_MINUTES = 75
 HEAD_TRAINING_KEEP_RUNS = 20
@@ -125,7 +121,7 @@ IMPORT_BATCH_KEEP_DAYS = 30
 # task.time_limit + a small buffer. task_name overrides take precedence
 # over queue overrides.
 #
-# ml queue: tag_and_embed video branch (≈20 GPU ops); time_limit=1200.
+# ml queue: embed_image video branch (≈20 GPU ops); time_limit=1200.
 # import_archive_file: shares the 'import' queue with the fast
 #   single-file import_media_file, so it needs a task-name override
 #   (the import queue itself stays at the 5-min default for single
@@ -143,10 +139,9 @@ QUEUE_STUCK_THRESHOLD_MINUTES: dict[str, int] = {
     "download": 30,
     # Audit 2026-06-02 — maintenance/scan queues run tasks that
     # legitimately exceed the 5-min default (verify_integrity at 70m
-    # hard, scan_directory at 70m hard, apply_allowlist_tags /
+    # hard, scan_directory at 70m hard, backfill_phash at 35m hard).
-    # recompute_centroids / backfill_phash at 35m hard). 75 min lives
+    # 75 min lives above the longest of those and the per-task
-    # above the longest of those and the per-task overrides below
+    # overrides below cover the outliers (backups, library audit).
-    # cover the outliers (backups, library audit).
     "maintenance": 75,
     "scan": 75,
 }
@@ -582,6 +577,28 @@ def verify_integrity() -> int:
     return total
 
 
+@celery.task(
+    name="backend.app.tasks.maintenance.triage_gpu_errors",
+    # Bounded small-set probe (only errored images, once each), but a single
+    # large original's sha256 over NFS can run tens of seconds — same quick-lane
+    # tolerance rationale as verify_integrity above.
+    soft_time_limit=600, time_limit=900,
+)
+def triage_gpu_errors() -> dict:
+    """Failure triage (#125): probe each errored GPU job's file once and write
+    the verdicts (ImageRecord.integrity_status + GpuJob.triage_status) — see
+    services/ml/gpu_triage.py. Time-boxed + resumable; no-op when every errored
+    job is already triaged."""
+    from ..services.ml.gpu_triage import triage_errored_jobs
+
+    SessionLocal = _sync_session_factory()
+    with SessionLocal() as session:
+        summary = triage_errored_jobs(session, time_budget_seconds=300.0)
+    if summary["probed"]:
+        log.info("triage_gpu_errors: %s", summary)
+    return summary
+
+
 @celery.task(name="backend.app.tasks.maintenance.recover_stalled_download_events")
 def recover_stalled_download_events() -> int:
     """Recover DownloadEvent rows stuck pending/running past the worker hard kill.
@@ -721,46 +738,6 @@ def recover_stalled_library_audit_runs() -> int:
         return recovered
 
 
-@celery.task(name="backend.app.tasks.maintenance.recover_stalled_tag_eval_runs")
-def recover_stalled_tag_eval_runs() -> int:
-    """Flip TagEvalRun rows stuck in 'running' past the stall threshold to
-    'error', and prune old runs to the last TAG_EVAL_KEEP_RUNS (retention,
-    rule 89). Runs every 5 min on the maintenance lane; no-op when idle."""
-    SessionLocal = _sync_session_factory()
-    now = datetime.now(UTC)
-    cutoff = now - timedelta(minutes=TAG_EVAL_STALL_THRESHOLD_MINUTES)
-    with SessionLocal() as session:
-        result = session.execute(
-            update(TagEvalRun)
-            .where(TagEvalRun.status == "running")
-            .where(
-                func.coalesce(TagEvalRun.last_progress_at, TagEvalRun.started_at)
-                < cutoff
-            )
-            .values(
-                status="error", finished_at=now,
-                error=(
-                    f"stranded by recovery sweep (no progress for "
-                    f"{TAG_EVAL_STALL_THRESHOLD_MINUTES} min)"
-                ),
-            )
-        )
-        # Retention: keep only the most recent N runs.
-        keep = session.execute(
-            select(TagEvalRun.id).order_by(TagEvalRun.id.desc())
-            .limit(TAG_EVAL_KEEP_RUNS)
-        ).scalars().all()
-        if keep:
-            session.execute(
-                delete(TagEvalRun).where(TagEvalRun.id.not_in(keep))
-            )
-        session.commit()
-        recovered = result.rowcount or 0
-        if recovered:
-            log.info("recover_stalled_tag_eval_runs: recovered %d rows", recovered)
-        return recovered
-
-
 @celery.task(name="backend.app.tasks.maintenance.recover_stalled_head_training_runs")
 def recover_stalled_head_training_runs() -> int:
     """Flip HeadTrainingRun rows stuck in 'running' past the stall threshold to

backend/app/tasks/ml.py

@@ -1,20 +1,26 @@
-"""ML Celery tasks: per-image inference, backfill discovery, centroid
+"""ML Celery tasks: per-image embedding, backfill discovery, head training,
-recompute, allowlist auto-apply, model self-heal.
+model self-heal.
 
-All run on the ml-worker (queue 'ml') except recompute_centroids and
+All run on the ml-worker (queue 'ml'), which under B3 (2026-07-02) is an
-apply_allowlist_tags sweeps which are 'maintenance' lane. Sync sessions
+OPTIONAL container: its only processing role is the CPU whole-image embed
-(Celery workers are sync processes), same pattern as FC-2a tasks.
+fallback (gated by ml_settings.cpu_embed_enabled) for stacks without a GPU
+agent — plus head training / auto-apply, which need sklearn/numpy and so
+live on this image. GPU-queue coordination (backfill enqueues, orphan
+recovery, reprocess) deliberately does NOT live here — see tasks/gpu_queue.py
+(maintenance lane), so the agent pipeline works with no ml-worker at all.
+Sync sessions (Celery workers are sync processes), same pattern as FC-2a
+tasks.
 """
 
 import logging
 from pathlib import Path
 
 from celery.exceptions import SoftTimeLimitExceeded
-from sqlalchemy import delete, select
+from sqlalchemy import select
 from sqlalchemy.exc import DBAPIError, OperationalError
 
 from ..celery_app import celery
-from ..models import ImagePrediction, ImageRecord, MLSettings
+from ..models import ImageRecord, MLSettings
 from ._sync_engine import sync_session_factory as _sync_session_factory
 
 log = logging.getLogger(__name__)
@@ -27,8 +33,24 @@ def _is_video(path: Path) -> bool:
     return path.suffix.lower() in VIDEO_EXTS
 
 
+def cpu_embed_enabled() -> bool:
+    """Dispatch gate for the CPU embed fallback (B3, operator 2026-07-02):
+    stacks that run a GPU agent and DROP the (optional) ml-worker container
+    turn ml_settings.cpu_embed_enabled off, so the import hooks stop queueing
+    embed work into a queue nothing consumes — the daily GPU 'embed' backfill
+    covers those images instead. Opens its own short session because the four
+    dispatch sites sit in different session scopes; defaults ON when the
+    settings row is missing (a fresh install must work agent-less)."""
+    SessionLocal = _sync_session_factory()
+    with SessionLocal() as session:
+        val = session.execute(
+            select(MLSettings.cpu_embed_enabled).where(MLSettings.id == 1)
+        ).scalar_one_or_none()
+    return True if val is None else bool(val)
+
+
 @celery.task(
-    name="backend.app.tasks.ml.tag_and_embed",
+    name="backend.app.tasks.ml.embed_image",
     bind=True,
     autoretry_for=(OperationalError, DBAPIError, OSError),
     retry_backoff=5,
@@ -45,20 +67,25 @@ def _is_video(path: Path) -> bool:
     soft_time_limit=900,   # 15 min
     time_limit=1200,       # 20 min hard
 )
-def tag_and_embed(self, image_id: int) -> dict:
+def embed_image(self, image_id: int) -> dict:
-    """Run Camie + SigLIP on one image; store predictions + embedding;
+    """Compute + store one image's whole-image SigLIP embedding — the CPU
-    then enqueue per-image allowlist application.
+    fallback path (B3, operator 2026-07-02): this is the ml-worker's ONLY
+    processing role, keeping search/similarity/head-suggestions alive on
+    deployments without a GPU agent. Detection, cropping and CCIP are
+    deliberately agent-only, and their backfill predicates read image_region /
+    gpu_job state — never image_record.siglip_embedding — so a CPU whole-image
+    embed can NEVER mark crop work as done. (Renamed from tag_and_embed —
+    Camie tagging was retired #1189; the old name kept implying a tagging step
+    that no longer exists.)
 
     Video (#747): sample frames at a fixed cadence (ml_settings
-    video_frame_interval_seconds, capped at video_max_frames), keep a tag only if
+    video_frame_interval_seconds, capped at video_max_frames) and mean-pool the
-    it appears in >= video_min_tag_frames frames and average its confidence over
+    per-frame SigLIP embeddings — the same shape as the GPU agent's video
-    those frames (mean-pool, not max — kills one-frame noise); mean-pool the
+    handling. On no-frames returns status='no_frames' (not an error).
-    SigLIP embeddings. On no-frames returns status='no_frames' (not an error).
     """
     import time
 
     from ..services.ml.embedder import get_embedder
-    from ..services.ml.tagger import get_tagger
 
     # Phase + file context, so a timeout/crash names WHICH file and WHERE it
     # died instead of a bare SoftTimeLimitExceeded() (operator-flagged 2026-06-08:
@@ -88,110 +115,64 @@ def tag_and_embed(self, image_id: int) -> dict:
                 f"image_id={image_id} path={record.path} mime={record.mime} "
                 f"bytes={record.size_bytes} video={is_vid}"
             )
-            log.info("tag_and_embed start: %s", ctx)
+            log.info("embed_image start: %s", ctx)
             if not src.is_file():
-                log.warning("tag_and_embed file missing on disk: %s", ctx)
+                log.warning("embed_image file missing on disk: %s", ctx)
                 return {"status": "file_missing", "image_id": image_id}
 
             phase = "load_models"
-            tagger = get_tagger()
+            embedder = get_embedder(settings.embedder_model_name)
-            embedder = get_embedder()
 
             if is_vid:
                 # Layer-3 isolation: ffprobe (a separate process) validates
-                # the container before we burn ~20 GPU ops sampling frames
+                # the container before we burn GPU ops sampling frames from it.
-                # from it. A corrupt video that would crash the frame
+                # A corrupt video that would crash the frame decoder is rejected
-                # decoder is rejected cleanly here instead of taking down
+                # cleanly here instead of taking down the ml-worker.
-                # the ml-worker. Operator-flagged 2026-05-28.
                 phase = "video_probe"
                 from ..utils import safe_probe
                 vprobe = safe_probe.probe_video(src)
                 if not vprobe.ok:
                     log.warning(
-                        "tag_and_embed bad video (%s): %s", vprobe.reason, ctx
+                        "embed_image bad video (%s): %s", vprobe.reason, ctx
                     )
                     return {
                         "status": "bad_video", "image_id": image_id,
                         "reason": vprobe.reason,
                     }
                 phase = "video_sample_frames"
-                t0 = time.monotonic()
                 frames = _sample_video_frames(
                     src,
                     interval=settings.video_frame_interval_seconds,
                     max_frames=settings.video_max_frames,
                 )
-                log.info(
-                    "tag_and_embed sampled %d frame(s) in %.1fs: %s",
-                    len(frames), time.monotonic() - t0, ctx,
-                )
                 if not frames:
                     return {"status": "no_frames", "image_id": image_id}
-                phase = "video_infer"
+                phase = "video_embed"
                 import numpy as np
 
-                preds = _aggregate_video_predictions(
+                # Mean-pool the per-frame SigLIP embeddings into one vector.
-                    [tagger.infer(f, store_floor=settings.tagger_store_floor)
-                     for f in frames],
-                    min_frames=settings.video_min_tag_frames,
-                )
                 embedding = np.mean(
                     [embedder.infer(f) for f in frames], axis=0
                 ).astype("float32")
-                log.info(
-                    "tag_and_embed video aggregated %d tag(s) from %d frame(s) "
-                    "(min_frames=%d): %s",
-                    len(preds), len(frames), settings.video_min_tag_frames, ctx,
-                )
                 for f in frames:
                     f.unlink(missing_ok=True)
             else:
-                phase = "tag"
-                t0 = time.monotonic()
-                raw = tagger.infer(src, store_floor=settings.tagger_store_floor)
-                log.info(
-                    "tag_and_embed tagged in %.1fs (%d tags): %s",
-                    time.monotonic() - t0, len(raw), ctx,
-                )
-                preds = {
-                    name: {"category": p.category, "confidence": p.confidence}
-                    for name, p in raw.items()
-                }
                 phase = "embed"
                 t0 = time.monotonic()
                 embedding = embedder.infer(src)
                 log.info(
-                    "tag_and_embed embedded in %.1fs: %s",
+                    "embed_image embedded in %.1fs: %s",
                     time.monotonic() - t0, ctx,
                 )
 
             phase = "persist"
-            record.tagger_model_version = settings.tagger_model_version
             record.siglip_embedding = embedding.tolist()
             record.siglip_model_version = settings.embedder_model_version
             session.add(record)
-            # Write the normalized image_prediction rows (#768) — the sole home
-            # for predictions now (image_record.tagger_predictions was dropped in
-            # migration 0046). Delete-then-insert keeps a re-tag idempotent;
-            # tagger_store_floor was already applied in tagger.infer, so preds is
-            # the >=floor set.
-            session.execute(
-                delete(ImagePrediction).where(
-                    ImagePrediction.image_record_id == image_id
-                )
-            )
-            session.add_all([
-                ImagePrediction(
-                    image_record_id=image_id, raw_name=name,
-                    category=p.get("category", "general"),
-                    score=float(p.get("confidence", 0.0)),
-                )
-                for name, p in preds.items()
-            ])
             session.commit()
     except SoftTimeLimitExceeded:
         log.error(
-            "tag_and_embed TIMED OUT after %.0fs in phase=%s: %s",
+            "embed_image TIMED OUT after %.0fs in phase=%s: %s",
             _elapsed(), phase, ctx,
         )
         # Re-raise as SoftTimeLimitExceeded (preserves the 'timeout' status in
@@ -205,16 +186,13 @@ def tag_and_embed(self, image_id: int) -> dict:
         # ORIGINAL so the type is preserved; just make sure it's logged with
         # context first.
         log.exception(
-            "tag_and_embed FAILED in phase=%s after %.0fs: %s",
+            "embed_image FAILED in phase=%s after %.0fs: %s",
             phase, _elapsed(), ctx,
         )
         raise
 
-    log.info(
+    log.info("embed_image ok in %.1fs: %s", _elapsed(), ctx)
-        "tag_and_embed ok in %.1fs (%d tags): %s", _elapsed(), len(preds), ctx
+    return {"status": "ok", "image_id": image_id}
-    )
-    apply_allowlist_tags.delay(image_id=image_id)
-    return {"status": "ok", "image_id": image_id, "tags": len(preds)}
 
 
 def _sample_video_frames(
@@ -273,318 +251,40 @@ def _sample_video_frames(
     return out
 
 
-def _aggregate_video_predictions(per_frame: list[dict], *, min_frames: int) -> dict:
-    """Aggregate per-frame {name: TagPrediction} into one prediction set (#747).
-
-    A tag is kept only if it appears (≥ the tagger store floor, already applied)
-    in at least `min_frames` of the sampled frames — because sampling is at a
-    fixed cadence, that means it was on screen for roughly min_frames×interval
-    seconds, so a single-frame flicker / scene-transition artifact is dropped
-    while a genuine scene-local tag in a long video survives. Confidence is the
-    MEAN over the frames where the tag appears (not max — max re-inflated the
-    one-frame noise this whole change exists to remove).
-
-    `min_frames` is clamped to the number of frames actually sampled so a very
-    short video (1–2 frames) still tags instead of dropping everything.
-    """
-    n = len(per_frame)
-    if n == 0:
-        return {}
-    threshold = max(1, min(min_frames, n))
-    agg: dict[str, dict] = {}
-    for frame_preds in per_frame:
-        for name, p in frame_preds.items():
-            cur = agg.get(name)
-            if cur is None:
-                agg[name] = {"category": p.category, "sum": p.confidence, "count": 1}
-            else:
-                cur["sum"] += p.confidence
-                cur["count"] += 1
-    return {
-        name: {"category": v["category"], "confidence": v["sum"] / v["count"]}
-        for name, v in agg.items()
-        if v["count"] >= threshold
-    }
-
-
 @celery.task(name="backend.app.tasks.ml.backfill", bind=True)
 def backfill(self) -> int:
-    """Enqueue tag_and_embed for images missing predictions/embeddings for
+    """Enqueue embed_image (embed-only) for images with no SigLIP embedding.
-    the current model versions. Keyset pagination by id ASC (restart-safe).
+    Keyset pagination by id ASC (restart-safe).
+
+    NB: a siglip MODEL-VERSION mismatch (an operator model swap, #1190) is NOT
+    re-embedded here — the CPU ml-worker can't churn the library at 384/512px;
+    the GPU agent owns version re-embeds via the 'embed' job.
     """
+    if not cpu_embed_enabled():
+        log.info("cpu backfill skipped: cpu_embed_enabled is off (B3 — the "
+                 "GPU 'embed' backfill owns whole-image embeds on this stack)")
+        return 0
     SessionLocal = _sync_session_factory()
     enqueued = 0
     last_id = 0
     with SessionLocal() as session:
-        settings = session.execute(
-            select(MLSettings).where(MLSettings.id == 1)
-        ).scalar_one()
         while True:
             rows = session.execute(
                 select(ImageRecord.id)
                 .where(ImageRecord.id > last_id)
-                .where(
+                .where(ImageRecord.siglip_embedding.is_(None))
-                    (ImageRecord.tagger_model_version.is_(None))
-                    | (
-                        ImageRecord.tagger_model_version
-                        != settings.tagger_model_version
-                    )
-                    | (ImageRecord.siglip_embedding.is_(None))
-                    | (
-                        ImageRecord.siglip_model_version
-                        != settings.embedder_model_version
-                    )
-                )
                 .order_by(ImageRecord.id.asc())
                 .limit(500)
             ).scalars().all()
             if not rows:
                 break
             for image_id in rows:
-                tag_and_embed.delay(image_id)
+                embed_image.delay(image_id)
                 enqueued += 1
             last_id = rows[-1]
     return enqueued
 
 
-@celery.task(
-    name="backend.app.tasks.ml.apply_allowlist_tags",
-    bind=True,
-    # Audit 2026-06-02 — the full-sweep mode (neither tag_id nor image_id)
-    # is O(images × allowlist) and legitimately runs >5 min on large
-    # libraries. Cap matches the maintenance queue's recovery threshold.
-    soft_time_limit=1800, time_limit=2100,
-)
-def apply_allowlist_tags(self, tag_id: int | None = None,
-                         image_id: int | None = None) -> int:
-    """Retroactively apply allowlisted tags.
-
-    Modes:
-    - tag_id only      : scan all images for this tag.
-    - image_id only    : scan all allowlisted tags for this image.
-    - both             : just the (image, tag) pair.
-    - neither          : full sweep (daily beat).
-
-    Skips: already-applied, rejected (tag_suggestion_rejection), or
-    confidence below the tag's allowlist min_confidence. Applied with
-    source='ml_auto'.
-    """
-    from sqlalchemy import and_
-    from sqlalchemy import select as sa_select
-    from sqlalchemy.dialects.postgresql import insert as pg_insert
-
-    from ..models import TagAllowlist, TagSuggestionRejection
-    from ..models.tag import image_tag
-
-    SessionLocal = _sync_session_factory()
-    applied = 0
-    with SessionLocal() as session:
-        allow_rows = session.execute(
-            sa_select(TagAllowlist.tag_id, TagAllowlist.min_confidence)
-            if tag_id is None
-            else sa_select(
-                TagAllowlist.tag_id, TagAllowlist.min_confidence
-            ).where(TagAllowlist.tag_id == tag_id)
-        ).all()
-        allow = {r[0]: r[1] for r in allow_rows}
-        if not allow:
-            return 0
-
-        # Images that have any predictions (#768: from image_prediction, not
-        # the old JSON column), optionally narrowed to one image.
-        img_ids_query = sa_select(ImagePrediction.image_record_id).distinct()
-        if image_id is not None:
-            img_ids_query = img_ids_query.where(
-                ImagePrediction.image_record_id == image_id
-            )
-
-        for (img_id,) in session.execute(img_ids_query).all():
-            preds = _load_predictions_sync(session, img_id)
-            for a_tag_id, min_conf in allow.items():
-                exists = session.execute(
-                    sa_select(image_tag.c.tag_id).where(
-                        and_(
-                            image_tag.c.image_record_id == img_id,
-                            image_tag.c.tag_id == a_tag_id,
-                        )
-                    )
-                ).scalar_one_or_none()
-                if exists is not None:
-                    continue
-                rej = session.get(
-                    TagSuggestionRejection, (img_id, a_tag_id)
-                )
-                if rej is not None:
-                    continue
-                from ..models import Tag
-
-                tag = session.get(Tag, a_tag_id)
-                if tag is None:
-                    continue
-                conf = _confidence_for_tag(session, tag, preds)
-                if conf is None or conf < min_conf:
-                    continue
-                stmt = pg_insert(image_tag).values(
-                    image_record_id=img_id,
-                    tag_id=a_tag_id,
-                    source="ml_auto",
-                )
-                stmt = stmt.on_conflict_do_nothing(
-                    index_elements=["image_record_id", "tag_id"]
-                )
-                session.execute(stmt)
-                applied += 1
-        session.commit()
-    return applied
-
-
-def _load_predictions_sync(session, image_id: int) -> dict:
-    """Predictions for one image from image_prediction (#768), in the
-    {raw_name: {category, confidence}} shape _confidence_for_tag consumes —
-    keeps the allowlist resolution logic unchanged."""
-    from sqlalchemy import select as sa_select
-
-    rows = session.execute(
-        sa_select(
-            ImagePrediction.raw_name,
-            ImagePrediction.category,
-            ImagePrediction.score,
-        ).where(ImagePrediction.image_record_id == image_id)
-    ).all()
-    return {
-        r.raw_name: {"category": r.category, "confidence": r.score}
-        for r in rows
-    }
-
-
-def _confidence_for_tag(session, tag, preds: dict) -> float | None:
-    """Highest confidence among predictions that resolve to `tag` —
-    either the prediction name equals the tag name, or an alias maps
-    (prediction name, category) -> tag.id.
-    """
-    from sqlalchemy import select as sa_select
-
-    from ..models import TagAlias
-
-    best: float | None = None
-    direct = preds.get(tag.name)
-    if direct is not None:
-        best = float(direct.get("confidence", 0.0))
-    alias_rows = session.execute(
-        sa_select(TagAlias.alias_string, TagAlias.alias_category).where(
-            TagAlias.canonical_tag_id == tag.id
-        )
-    ).all()
-    for alias_string, alias_category in alias_rows:
-        p = preds.get(alias_string)
-        if p is None:
-            continue
-        if p.get("category") != alias_category:
-            continue
-        c = float(p.get("confidence", 0.0))
-        if best is None or c > best:
-            best = c
-    return best
-
-
-@celery.task(name="backend.app.tasks.ml.recompute_centroid", bind=True)
-def recompute_centroid(self, tag_id: int) -> bool:
-    import asyncio
-
-    from ..services.ml.centroids import CentroidService
-    from ._async_session import async_session_factory
-
-    async def _run() -> bool:
-        # Per-task NullPool engine bound to THIS asyncio.run loop — the shared
-        # process-wide engine reuses connections across loops and raises
-        # "Future attached to a different loop" on every call after the first.
-        async_factory, async_engine = async_session_factory()
-        try:
-            async with async_factory() as session:
-                svc = CentroidService(session)
-                result = await svc.recompute_for_tag(tag_id)
-                await session.commit()
-                return result
-        finally:
-            await async_engine.dispose()
-
-    return asyncio.run(_run())
-
-
-@celery.task(
-    name="backend.app.tasks.ml.recompute_centroids",
-    bind=True,
-    # Audit 2026-06-02 — drifted-centroid rebuild over potentially
-    # hundreds of tags.
-    soft_time_limit=1800, time_limit=2100,
-)
-def recompute_centroids(self) -> int:
-    """Daily: find drifted centroids, enqueue recompute_centroid for each."""
-    import asyncio
-
-    from ..services.ml.centroids import CentroidService
-    from ._async_session import async_session_factory
-
-    async def _list() -> list[int]:
-        # Per-task NullPool engine bound to this loop (see recompute_centroid).
-        async_factory, async_engine = async_session_factory()
-        try:
-            async with async_factory() as session:
-                return await CentroidService(session).list_drifted()
-        finally:
-            await async_engine.dispose()
-
-    drifted = asyncio.run(_list())
-    for tid in drifted:
-        recompute_centroid.delay(tid)
-    return len(drifted)
-
-
-@celery.task(
-    name="backend.app.tasks.ml.tag_eval_run",
-    bind=True,
-    # The head-vs-centroid eval (#1130) loads embeddings + fits sklearn heads
-    # for several concepts — minutes, not seconds. Runs on the ml queue because
-    # only that worker has numpy/scikit-learn.
-    soft_time_limit=1800, time_limit=2100,
-)
-def tag_eval_run(self, run_id: int) -> str:
-    """Compute the eval report into the persisted TagEvalRun row so it survives
-    navigation (the admin card rehydrates from the row, not transient state)."""
-    from datetime import UTC, datetime
-
-    from ..models import TagEvalRun
-    from ..services.ml.tag_eval import run_eval
-
-    SessionLocal = _sync_session_factory()
-    with SessionLocal() as session:
-        run = session.get(TagEvalRun, run_id)
-        if run is None:
-            return "missing"
-        run.last_progress_at = datetime.now(UTC)
-        session.commit()
-        try:
-            report = run_eval(session, run.params)
-        except SoftTimeLimitExceeded:
-            run.status = "error"
-            run.error = "timed out"
-            run.finished_at = datetime.now(UTC)
-            session.commit()
-            raise
-        except Exception as exc:
-            log.exception("tag_eval_run %d failed", run_id)
-            run.status = "error"
-            run.error = str(exc)
-            run.finished_at = datetime.now(UTC)
-            session.commit()
-            return "error"
-        run.report = report
-        run.status = "ready"
-        run.finished_at = datetime.now(UTC)
-        session.commit()
-        return "ready"
-
-
 @celery.task(
     name="backend.app.tasks.ml.train_heads",
     bind=True,
@@ -740,82 +440,6 @@ def scheduled_apply_head_tags() -> str:
     return "dispatched"
 
 
-@celery.task(name="backend.app.tasks.ml.enqueue_gpu_backfill")
-def enqueue_gpu_backfill(task_name: str) -> int:
-    """Enqueue a gpu_job for every image that still needs `task_name` (one
-    INSERT…SELECT, so it scales to a full library). The desktop agent drains the
-    queue over HTTP. Returns the number enqueued.
-
-    'siglip' gates on the RESULT (no concept region yet) rather than on a prior
-    job, so it picks up the back-catalogue of images that were CCIP-embedded
-    before concept crops existed, and retries images whose concept embed failed —
-    without re-touching their figure/CCIP regions."""
-    from sqlalchemy import exists, insert, literal
-    from sqlalchemy import select as sa_select
-
-    from ..models import GpuJob, ImageRecord, ImageRegion
-
-    SessionLocal = _sync_session_factory()
-    with SessionLocal() as session:
-        if task_name == "siglip":
-            has_concept = exists().where(
-                ImageRegion.image_record_id == ImageRecord.id,
-                ImageRegion.kind == "concept",
-            )
-            queued = exists().where(
-                GpuJob.image_record_id == ImageRecord.id,
-                GpuJob.task == "siglip",
-                GpuJob.status.in_(["pending", "leased"]),
-            )
-            sel = sa_select(
-                ImageRecord.id, literal("siglip"), literal("pending")
-            ).where(~has_concept).where(~queued)
-        else:
-            already = exists().where(
-                GpuJob.image_record_id == ImageRecord.id,
-                GpuJob.task == task_name,
-                GpuJob.status.in_(["pending", "leased", "done"]),
-            )
-            sel = sa_select(
-                ImageRecord.id, literal(task_name), literal("pending")
-            ).where(~already)
-        # RETURNING + count: result.rowcount is unreliable for INSERT…SELECT.
-        rows = session.execute(
-            insert(GpuJob)
-            .from_select(["image_record_id", "task", "status"], sel)
-            .returning(GpuJob.id)
-        ).fetchall()
-        session.commit()
-        return len(rows)
-
-
-@celery.task(name="backend.app.tasks.ml.recover_orphaned_gpu_jobs")
-def recover_orphaned_gpu_jobs() -> int:
-    """Reset expired GPU-job leases back to pending — recovers work orphaned by an
-    agent that died mid-job (no graceful release). Short beat cadence so orphans
-    get picked back up quickly + the queue counts read honestly. Returns the
-    number recovered."""
-    from datetime import UTC, datetime
-
-    from sqlalchemy import update
-
-    from ..models import GpuJob
-
-    SessionLocal = _sync_session_factory()
-    with SessionLocal() as session:
-        now = datetime.now(UTC)
-        res = session.execute(
-            update(GpuJob)
-            .where(GpuJob.status == "leased", GpuJob.lease_expires_at < now)
-            .values(
-                status="pending", lease_token=None, leased_at=None,
-                lease_expires_at=None, updated_at=now,
-            )
-        )
-        session.commit()
-        return res.rowcount or 0
-
-
 @celery.task(
     name="backend.app.tasks.ml.scheduled_ccip_auto_apply",
     soft_time_limit=1800, time_limit=2100,

entrypoint.sh

@@ -12,9 +12,13 @@ case "$ROLE" in
     # create_app is a factory — the `()` tells hypercorn to call it once
     # and serve the returned Quart (ASGI) app, rather than treating the
     # function itself as the application (which it then mis-invokes as WSGI).
+    # Default 4 workers (was 2): each worker is one asyncio loop, and a large
+    # file download occupies its worker for the transfer — 2 was too few once the
+    # GPU agent + the browser's thumbnail grid hit /images concurrently (they
+    # queued behind each other). Env-tunable via HYPERCORN_WORKERS.
     exec hypercorn \
       --bind 0.0.0.0:8080 \
-      --workers "${HYPERCORN_WORKERS:-2}" \
+      --workers "${HYPERCORN_WORKERS:-4}" \
       --access-logfile - \
       "backend.app:create_app()"
     ;;

frontend/src/components/common/MaintenanceTile.vue

@@ -7,8 +7,10 @@
 
   Usage: wrap a card's action body in the default slot; pass icon/title/blurb.
   `destructive` tints the icon error-red for delete actions. `open` can be forced
-  (e.g. keep a running task's tile expanded). Keyboard accessible: the header is a
+  (e.g. keep a running task's tile expanded). Manual expand/collapse persists per
-  real <button> with aria-expanded + focus ring.
+  tile in localStorage, so the page reloads the way the operator left it.
+  Keyboard accessible: the header is a real <button> with aria-expanded + focus
+  ring.
 -->
 <template>
   <v-card class="fc-tile" :class="{ 'fc-tile--open': isOpen }">
@@ -53,12 +55,21 @@ const props = defineProps({
   open: { type: Boolean, default: false },
 })
 
-const local = ref(props.open)
+// Only MANUAL toggles are saved (keyed by tile title): a forced `open` while a
-watch(() => props.open, (v) => { local.value = v })
+// task is mid-run is transient state, not a preference — persisting it would
+// resurrect the "several tiles open by default" bug this replaces. When the
+// force clears, the tile falls back to the operator's saved choice.
+const storeKey = `fc.tile.${props.title}`
+function savedOpen() {
+  try { return localStorage.getItem(storeKey) === '1' } catch { return false }
+}
+const local = ref(props.open || savedOpen())
+watch(() => props.open, (v) => { local.value = v || savedOpen() })
 const isOpen = computed(() => local.value)
 
 function toggle() {
   local.value = !local.value
+  try { localStorage.setItem(storeKey, local.value ? '1' : '0') } catch { /* non-fatal */ }
 }
 </script>
 

frontend/src/components/gallery/GalleryFilterBar.vue

@@ -142,9 +142,13 @@ const tagStore = useTagStore()
 const api = useApi()
 const router = useRouter()
 
+// posted_* sort by earliest publish across ALL of an image's posts (original
+// post date); newest/oldest sort by the primary post's date, else download date.
 const SORTS = [
-  { title: 'Newest first', value: 'newest' },
+  { title: 'Newest post date', value: 'posted_new' },
-  { title: 'Oldest first', value: 'oldest' },
+  { title: 'Oldest post date', value: 'posted_old' },
+  { title: 'Newest added', value: 'newest' },
+  { title: 'Oldest added', value: 'oldest' },
 ]
 
 const selected = ref(null)
@@ -175,7 +179,7 @@ const hasActiveFilters = computed(() =>
   store.filter.tag_or.length > 0 ||
   store.filter.artist_id != null ||
   store.filter.media_type != null ||
-  store.filter.sort !== 'newest' ||
+  store.filter.sort !== 'posted_new' ||
   store.filter.similar_to != null ||
   hasRefineFilters.value
 )

frontend/src/components/modal/SuggestionItem.vue

@@ -132,8 +132,8 @@ const hasMenu = computed(() =>
   color: rgb(var(--v-theme-on-surface-variant, var(--v-theme-on-surface)));
   font-family: 'JetBrains Mono', monospace;
 }
-/* Green ✓ / red ✗ verdict pair — same circular language as the eval card
+/* Green ✓ / red ✗ verdict pair — circular buttons so accept/reject read
-   (TagEvalCard .fc-act) so accept/reject read identically across surfaces. */
+   identically across surfaces. */
 .fc-suggestion__acts {
   flex: 0 0 auto; display: flex; gap: 4px;
 }

frontend/src/components/settings/AllowlistTable.vue

@@ -1,120 +0,0 @@
-<template>
-  <MaintenanceTile
-    icon="mdi-playlist-check"
-    :title="`Allowlisted tags (${store.rows.length})`"
-    blurb="Tags auto-applied to images that score above their threshold. Tune the
-           threshold and see how many images it would cover."
-  >
-    <v-data-table-virtual
-      :headers="headers" :items="store.rows" :loading="store.loading"
-      height="360" density="compact" fixed-header
-      no-data-text="No tags on the allowlist yet — accept a suggestion to add one."
-    >
-      <template #item.applied_count="{ item }">
-        <span class="fc-num">{{ item.applied_count ?? '—' }}</span>
-      </template>
-
-      <template #item.min_confidence="{ item }">
-        <div class="fc-thr">
-          <v-text-field
-            :model-value="item.min_confidence" type="number"
-            density="compact" hide-details style="max-width: 100px;"
-            :min="floor" max="1" step="0.05"
-            :aria-label="`Auto-apply threshold for ${item.tag_name}`"
-            @update:model-value="(v) => onThreshold(item, v)"
-          />
-          <span
-            v-if="proj[item.tag_id]"
-            class="fc-thr__proj"
-            :class="{ 'fc-thr__proj--loading': proj[item.tag_id].loading }"
-            :title="`At ${proj[item.tag_id].threshold}, a sweep would cover this many images`"
-          >≈ {{ proj[item.tag_id].count }} at {{ proj[item.tag_id].threshold }}</span>
-        </div>
-      </template>
-
-      <template #item.coverage_count="{ item }">
-        <span class="fc-num" :title="`Images a sweep covers at ${item.min_confidence}`">
-          {{ item.coverage_count ?? '—' }}
-        </span>
-      </template>
-
-      <template #item.actions="{ item }">
-        <v-btn
-          icon="mdi-delete" size="x-small" variant="text" color="error"
-          :aria-label="`Remove ${item.tag_name} from the allowlist`"
-          @click="store.remove(item.tag_id)"
-        />
-      </template>
-    </v-data-table-virtual>
-    <p class="fc-muted text-caption mt-2">
-      <strong>Applied</strong> = images currently carrying the tag.
-      <strong>Covers</strong> = images a sweep would auto-apply it to at the
-      current threshold. Lower the threshold to cover more (less certain) images.
-    </p>
-  </MaintenanceTile>
-</template>
-
-<script setup>
-import { computed, onMounted, reactive } from 'vue'
-import { useAllowlistStore } from '../../stores/allowlist.js'
-import { useMLStore } from '../../stores/ml.js'
-import MaintenanceTile from '../common/MaintenanceTile.vue'
-
-const store = useAllowlistStore()
-const ml = useMLStore()
-// min_confidence can't be set below the tagger store floor — predictions
-// below it aren't stored, so a lower threshold would behave identically to
-// the floor. The backend clamps too (#764).
-const floor = computed(() => ml.settings?.tagger_store_floor ?? 0.70)
-const headers = [
-  { title: 'Tag', key: 'tag_name', sortable: true },
-  { title: 'Kind', key: 'tag_kind', sortable: true, width: 100 },
-  { title: 'Applied', key: 'applied_count', sortable: true, width: 90 },
-  { title: 'Min confidence', key: 'min_confidence', sortable: false, width: 220 },
-  { title: 'Covers', key: 'coverage_count', sortable: true, width: 90 },
-  { title: '', key: 'actions', sortable: false, width: 56 }
-]
-
-// Per-row live projection while the operator drags a threshold:
-//   proj[tagId] = { threshold, count, loading }
-const proj = reactive({})
-
-onMounted(() => {
-  store.load()
-  if (!ml.settings) ml.loadSettings()
-})
-
-const debounces = {}
-function onThreshold(item, value) {
-  const tagId = item.tag_id
-  const v = Math.max(parseFloat(value), floor.value)
-  if (!(v > 0 && v <= 1)) return
-  const shown = Number(v.toFixed(2))
-  // Optimistic live projection box (loading until the count returns).
-  proj[tagId] = { threshold: shown, count: proj[tagId]?.count ?? '…', loading: true }
-  if (debounces[tagId]) clearTimeout(debounces[tagId])
-  debounces[tagId] = setTimeout(async () => {
-    try {
-      const { count } = await store.coverage(tagId, v)
-      proj[tagId] = { threshold: shown, count, loading: false }
-    } catch {
-      delete proj[tagId]  // drop the projection rather than show a wrong number
-    }
-    // Commit the new threshold (also refreshes the row's stored coverage_count).
-    store.updateThreshold(tagId, v)
-  }, 500)
-}
-</script>
-
-<style scoped>
-.fc-num { font-variant-numeric: tabular-nums; }
-.fc-thr { display: flex; align-items: center; gap: 10px; }
-.fc-thr__proj {
-  font-size: 12px;
-  font-variant-numeric: tabular-nums;
-  color: rgb(var(--v-theme-accent));
-  white-space: nowrap;
-}
-.fc-thr__proj--loading { color: rgb(var(--v-theme-on-surface-variant)); }
-.fc-muted { color: rgb(var(--v-theme-on-surface-variant)); }
-</style>

frontend/src/components/settings/CentroidRecomputeCard.vue

@@ -1,36 +0,0 @@
-<template>
-  <MaintenanceTile
-    icon="mdi-vector-triangle"
-    title="Tag centroids"
-    blurb="Rebuild SigLIP centroids for similarity suggestions."
-    :open="busy"
-  >
-      <p class="text-body-2 mb-3">
-        Rebuild the per-tag SigLIP centroids that power similarity-based
-        suggestions. Runs nightly automatically; trigger manually after a
-        large tagging session.
-      </p>
-      <v-btn color="primary" rounded="pill" :loading="busy" @click="run">
-        <v-icon start>mdi-vector-triangle</v-icon> Recompute centroids
-      </v-btn>
-      <span v-if="done" class="ml-3 text-caption">Enqueued.</span>
-      <QueueStatusBar queue="ml" queue-label="ML" />
-  </MaintenanceTile>
-</template>
-
-<script setup>
-import { toast } from '../../utils/toast.js'
-import { ref } from 'vue'
-import { useMLStore } from '../../stores/ml.js'
-import MaintenanceTile from '../common/MaintenanceTile.vue'
-import QueueStatusBar from './QueueStatusBar.vue'
-const store = useMLStore()
-const busy = ref(false)
-const done = ref(false)
-async function run() {
-  busy.value = true
-  try { await store.triggerRecomputeCentroids(); done.value = true }
-  catch (e) { toast({ text: e.message, type: 'error' }) }
-  finally { busy.value = false }
-}
-</script>

frontend/src/components/settings/DangerZoneCard.vue

@@ -0,0 +1,122 @@
+<template>
+  <MaintenanceTile
+    icon="mdi-nuke"
+    title="Reset content tagging (whole instance)"
+    blurb="Delete ALL general/character tags and their applications — a start-over. Requires a confirmation code."
+    destructive
+  >
+    <p class="text-body-2 mb-2">
+      Deletes every <code>general</code> and <code>character</code> tag and
+      removes them from every image — <strong>including the examples the
+      tagging heads learned from</strong>. Suggestions will <strong>not</strong>
+      repopulate on their own: you re-tag from scratch, and the heads retrain
+      from your new tags as they accumulate. Fandoms and series (with their
+      page order) are kept.
+    </p>
+    <v-alert type="error" variant="tonal" density="compact" class="mb-3">
+      Irreversible — no undo except restoring a DB backup
+      (Settings → Maintenance → Backup). Back one up first.
+    </v-alert>
+
+    <v-btn
+      color="accent" variant="flat" rounded="pill"
+      prepend-icon="mdi-magnify"
+      :loading="loadingPreview"
+      class="mb-3"
+      @click="onPreview"
+    >Preview content-tag reset</v-btn>
+
+    <div v-if="preview">
+      <p class="text-body-2 mb-2">
+        <strong>{{ preview.count }}</strong> content tag(s)
+        <span v-for="(n, k) in preview.by_kind" :key="k" class="fc-muted">
+          ({{ k }}: {{ n }})&nbsp;
+        </span>
+        across <strong>{{ preview.applications }}</strong> image
+        application(s).
+      </p>
+      <SampleNameGrid
+        v-if="preview.sample_names?.length"
+        :names="preview.sample_names" class="mb-3"
+      />
+      <template v-if="preview.count">
+        <p class="text-body-2 mb-2">
+          To arm the reset, type the confirmation code
+          <code class="fc-code">{{ preview.confirm }}</code> below.
+        </p>
+        <div class="d-flex align-center mb-1" style="gap: 12px">
+          <v-text-field
+            v-model="typed" density="compact" hide-details variant="outlined"
+            label="Confirmation code" style="max-width: 200px"
+            autocomplete="off" spellcheck="false"
+          />
+          <v-btn
+            color="error" variant="flat" rounded="pill"
+            prepend-icon="mdi-delete-alert"
+            :disabled="typed !== preview.confirm"
+            :loading="committing"
+            @click="onCommit"
+          >Delete {{ preview.count }} tag(s) +
+            {{ preview.applications }} application(s)</v-btn>
+        </div>
+        <p class="fc-muted text-caption mb-0">
+          The code is derived from the counts above — if tagging changes
+          between preview and apply, the server rejects the stale code.
+        </p>
+      </template>
+    </div>
+  </MaintenanceTile>
+</template>
+
+<script setup>
+import { ref } from 'vue'
+
+import { toast } from '../../utils/toast.js'
+import MaintenanceTile from '../common/MaintenanceTile.vue'
+import SampleNameGrid from '../common/SampleNameGrid.vue'
+import { useAdminStore } from '../../stores/admin.js'
+
+const store = useAdminStore()
+const preview = ref(null)
+const loadingPreview = ref(false)
+const committing = ref(false)
+const typed = ref('')
+
+async function onPreview() {
+  loadingPreview.value = true
+  typed.value = ''
+  try {
+    preview.value = await store.resetContentTagging({ dryRun: true })
+  } catch (e) {
+    toast({ text: `Preview failed: ${e.message}`, type: 'error' })
+  } finally {
+    loadingPreview.value = false
+  }
+}
+
+async function onCommit() {
+  committing.value = true
+  try {
+    const res = await store.resetContentTagging({
+      dryRun: false, confirm: typed.value,
+    })
+    toast({ text: `Deleted ${res.deleted} content tag(s) — re-tagging starts fresh`, type: 'success' })
+    preview.value = null
+    typed.value = ''
+  } catch (e) {
+    toast({ text: `Reset rejected: ${e.message}`, type: 'error' })
+  } finally {
+    committing.value = false
+  }
+}
+</script>
+
+<style scoped>
+.fc-muted { color: rgb(var(--v-theme-on-surface-variant)); }
+.fc-code {
+  background: rgb(var(--v-theme-surface-light));
+  border-radius: 4px; padding: 2px 8px;
+  font-family: 'JetBrains Mono', monospace; font-weight: 700;
+  letter-spacing: 0.06em;
+}
+</style>

frontend/src/components/settings/DownloadsActivityPanel.vue

@@ -0,0 +1,77 @@
+<template>
+  <v-card class="mb-4">
+    <CardHeading icon="mdi-download" title="Downloads (last 24h)">
+      <v-spacer />
+      <v-btn
+        variant="text" size="small" rounded="pill"
+        to="/subscriptions?tab=downloads"
+      >
+        Open subscriptions
+        <v-icon end size="small">mdi-arrow-right</v-icon>
+      </v-btn>
+    </CardHeading>
+    <v-card-text>
+      <div class="d-flex align-center flex-wrap mb-2" style="gap: 8px">
+        <v-chip size="small" variant="tonal" color="success">
+          {{ stats.ok }} ok
+        </v-chip>
+        <v-chip
+          size="small" variant="tonal"
+          :color="stats.error ? 'error' : undefined"
+        >
+          {{ stats.error }} failed
+        </v-chip>
+        <v-chip v-if="stats.running" size="small" variant="tonal" color="accent">
+          {{ stats.running }} running
+        </v-chip>
+        <v-chip v-if="stats.pending" size="small" variant="tonal">
+          {{ stats.pending }} pending
+        </v-chip>
+        <v-chip v-if="stats.skipped" size="small" variant="tonal">
+          {{ stats.skipped }} skipped
+        </v-chip>
+      </div>
+      <p v-if="!failing.length" class="fc-muted text-body-2 mb-0">
+        All subscription sources healthy.
+      </p>
+      <p v-else class="text-body-2 mb-0">
+        <b class="fc-bad">{{ failing.length }}</b> failing source(s):
+        <span class="fc-muted">{{ failingNames }}</span>
+      </p>
+    </v-card-text>
+  </v-card>
+</template>
+
+<script setup>
+import { computed, onMounted, onUnmounted } from 'vue'
+import { storeToRefs } from 'pinia'
+
+import CardHeading from '../common/CardHeading.vue'
+import { useDownloadsStore } from '../../stores/downloads.js'
+
+const store = useDownloadsStore()
+const { stats, failing } = storeToRefs(store)
+let pollId = null
+
+const failingNames = computed(() => {
+  const names = failing.value.map((s) => s.artist_name || s.url).slice(0, 3)
+  const extra = failing.value.length - names.length
+  return names.join(', ') + (extra > 0 ? ` +${extra} more` : '')
+})
+
+function poll() {
+  store.loadStats(24)
+  store.loadFailing()
+}
+
+onMounted(() => {
+  poll()
+  pollId = setInterval(() => { if (!document.hidden) poll() }, 30000)
+})
+onUnmounted(() => { if (pollId) clearInterval(pollId) })
+</script>
+
+<style scoped>
+.fc-muted { color: rgb(var(--v-theme-on-surface-variant)); }
+.fc-bad { color: rgb(var(--v-theme-error)); }
+</style>

frontend/src/components/settings/GpuActivityPanel.vue

@@ -0,0 +1,110 @@
+<template>
+  <v-card class="mb-4">
+    <CardHeading icon="mdi-expansion-card" title="GPU agent pipeline">
+      <v-spacer />
+      <v-btn
+        variant="text" size="small" rounded="pill"
+        @click="$emit('open-maintenance')"
+      >
+        Open maintenance
+        <v-icon end size="small">mdi-arrow-right</v-icon>
+      </v-btn>
+    </CardHeading>
+    <v-card-text>
+      <div class="fc-cells mb-2">
+        <div class="fc-cell">
+          <div class="fc-cell__n">{{ q.pending }}</div>
+          <div class="fc-cell__l">pending</div>
+        </div>
+        <div class="fc-cell">
+          <div class="fc-cell__n">{{ q.leased }}</div>
+          <div class="fc-cell__l">in flight</div>
+        </div>
+        <div class="fc-cell">
+          <div class="fc-cell__n fc-good">{{ q.done }}</div>
+          <div class="fc-cell__l">done</div>
+        </div>
+        <div class="fc-cell">
+          <div class="fc-cell__n" :class="q.error ? 'fc-bad' : ''">{{ q.error }}</div>
+          <div class="fc-cell__l">errored</div>
+        </div>
+      </div>
+      <p v-if="!q.error" class="fc-muted text-body-2 mb-0">
+        No failed jobs — the pipeline is clean. Work drains whenever the
+        desktop agent is running.
+      </p>
+      <p v-else class="text-body-2 mb-0">
+        Triage: <b>{{ triage.defect }}</b> defective file(s) ·
+        {{ triage.file_ok }} file-ok · {{ triage.unclassified }} unprobed
+        <span v-if="reasonSummary" class="fc-muted">— {{ reasonSummary }}</span>
+        <br>
+        <span class="fc-muted text-caption">
+          Recover defective files from Maintenance → Failed processing.
+        </span>
+      </p>
+    </v-card-text>
+  </v-card>
+</template>
+
+<script setup>
+import { computed, onMounted, onUnmounted, ref } from 'vue'
+
+import CardHeading from '../common/CardHeading.vue'
+import { useGpuStore } from '../../stores/gpu.js'
+
+defineEmits(['open-maintenance'])
+
+const store = useGpuStore()
+const q = ref({ pending: 0, leased: 0, done: 0, error: 0 })
+const triage = ref({ defect: 0, file_ok: 0, unclassified: 0 })
+const byClass = ref({})
+let pollId = null
+let lastErrorCount = -1
+
+const reasonSummary = computed(() =>
+  Object.entries(byClass.value)
+    .sort((a, b) => b[1] - a[1])
+    .slice(0, 3)
+    .map(([k, n]) => `${k.replaceAll('_', ' ')} ${n}`)
+    .join(' · '))
+
+async function poll() {
+  try {
+    q.value = await store.status()
+    // The triage detail is only worth a second call when the error count
+    // actually moved (it's a 500-row join server-side).
+    if (q.value.error !== lastErrorCount) {
+      lastErrorCount = q.value.error
+      if (q.value.error > 0) {
+        const body = await store.errors()
+        triage.value = body.triage
+        byClass.value = body.by_class
+      } else {
+        triage.value = { defect: 0, file_ok: 0, unclassified: 0 }
+        byClass.value = {}
+      }
+    }
+  } catch { /* non-fatal — panel just shows the last snapshot */ }
+}
+
+onMounted(() => {
+  poll()
+  pollId = setInterval(() => { if (!document.hidden) poll() }, 5000)
+})
+onUnmounted(() => { if (pollId) clearInterval(pollId) })
+</script>
+
+<style scoped>
+.fc-muted { color: rgb(var(--v-theme-on-surface-variant)); }
+.fc-cells { display: flex; gap: 28px; }
+.fc-cell__n {
+  font-size: 20px; font-weight: 700; line-height: 1.1;
+  font-family: 'JetBrains Mono', monospace;
+}
+.fc-cell__l {
+  font-size: 11px; text-transform: uppercase; letter-spacing: 0.04em;
+  color: rgb(var(--v-theme-on-surface-variant));
+}
+.fc-good { color: rgb(var(--v-theme-success)); }
+.fc-bad { color: rgb(var(--v-theme-error)); }
+</style>

frontend/src/components/settings/GpuAgentCard.vue

@@ -3,7 +3,6 @@
     icon="mdi-expansion-card"
     title="GPU agent (CCIP + crops)"
     blurb="Connect a desktop-GPU agent to embed characters (CCIP) and crops. It pulls work over HTTP — your database and Redis stay private."
-    :open="true"
   >
     <p class="fc-muted text-body-2 mb-3">
       The agent is a container you run on the machine with the GPU. It
@@ -52,6 +51,17 @@
       <div class="fc-q"><div class="fc-q__n" :class="queue.error ? 'fc-weak' : ''">{{ queue.error }}</div><div class="fc-q__l">errored</div></div>
     </div>
 
+    <v-btn
+      class="mt-3" color="accent" variant="tonal" rounded="pill" size="small"
+      prepend-icon="mdi-restart-alert" :loading="retrying"
+      :disabled="!queue.error" @click="onRetryErrors"
+    >Retry errored jobs</v-btn>
+    <p class="fc-muted text-caption mt-2 mb-0">
+      Errored jobs park after 3 failed attempts. This requeues just those (their
+      errors cleared, attempts reset) — use after updating the agent, without
+      re-running the whole done library.
+    </p>
+
     <v-btn
       class="mt-4" color="accent" variant="tonal" rounded="pill" size="small"
       prepend-icon="mdi-account-box-multiple" :loading="backfilling" @click="onBackfill"
@@ -71,6 +81,16 @@
       images get these automatically; this catches the back-catalogue.
     </p>
 
+    <v-btn
+      class="mt-3" color="warning" variant="tonal" rounded="pill" size="small"
+      prepend-icon="mdi-backup-restore" :loading="reprocessing" @click="onReprocess"
+    >Re-process library (re-detect + re-crop)</v-btn>
+    <p class="fc-muted text-caption mt-2 mb-0">
+      Re-runs the FULL pipeline (figure detection + CCIP + concept/panel crops) on
+      <b>every</b> image — use after changing crop detectors so the back-catalogue
+      gets re-cropped, not just new images. Heavy: re-processes the whole library.
+    </p>
+
     <!-- Match strictness -->
     <div class="fc-section-h mt-5 mb-1">Character-match strictness</div>
     <div v-if="ml.settings" class="d-flex align-center" style="gap:12px">
@@ -106,6 +126,33 @@
       reversible) — so identity tags keep flowing without review. Stricter than
       the suggest cut; 0.92 recommended.
     </p>
+
+    <!-- Embedding model -->
+    <div v-if="ml.settings" class="fc-section-h mt-5 mb-1">Embedding model</div>
+    <div v-if="ml.settings">
+      <v-select
+        v-model="selectedModel" :items="modelItems" item-title="label"
+        item-value="name" label="Model" density="compact" hide-details
+        variant="outlined"
+      />
+      <div class="d-flex mt-3" style="gap:8px">
+        <v-btn
+          size="small" variant="tonal" rounded="pill" :loading="savingModel"
+          prepend-icon="mdi-content-save" @click="onSaveModel"
+        >Save model</v-btn>
+        <v-btn
+          size="small" color="accent" variant="flat" rounded="pill"
+          :loading="reembedding" prepend-icon="mdi-backup-restore" @click="onReembed"
+        >Re-embed library (GPU)</v-btn>
+      </div>
+      <p class="fc-muted text-caption mt-2 mb-0">
+        Switching the model is a DIFFERENT embedding space. After <b>Save model</b>,
+        run <b>Re-embed library</b> (the GPU agent re-embeds whole images + concept
+        crops), then <b>Retrain heads</b> — suggestions degrade until both finish.
+        SigLIP 2 (512px) is a 1152-d drop-in over SigLIP 1; new installs default to
+        it. Your existing library stays on its current model until you re-embed.
+      </p>
+    </div>
   </MaintenanceTile>
 </template>
 
@@ -126,11 +173,17 @@ const masked = ref(true)
 const rotating = ref(false)
 const backfilling = ref(false)
 const backfillingSiglip = ref(false)
+const reprocessing = ref(false)
+const retrying = ref(false)
 const threshold = ref(0.85)
 const savingThreshold = ref(false)
 const autoApply = ref(true)
 const autoThreshold = ref(0.92)
 const savingAuto = ref(false)
+const modelItems = ref([])
+const selectedModel = ref(null)
+const savingModel = ref(false)
+const reembedding = ref(false)
 const queue = ref({ pending: 0, leased: 0, done: 0, error: 0 })
 let pollTimer = null
 
@@ -157,9 +210,50 @@ onMounted(async () => {
       autoApply.value = ml.settings.ccip_auto_apply_enabled
       autoThreshold.value = ml.settings.ccip_auto_apply_threshold
     }
+    if (ml.settings?.embedder_model_name != null) {
+      const items = await ml.embedderModels()
+      // Make sure the current model is selectable even if it's not in the list.
+      const cur = ml.settings.embedder_model_name
+      if (!items.some((m) => m.name === cur)) {
+        items.push({ name: cur, version: ml.settings.embedder_model_version, label: `${cur} (current)` })
+      }
+      modelItems.value = items
+      selectedModel.value = cur
+    }
   } catch { /* non-fatal */ }
 })
 
+async function onSaveModel() {
+  const opt = modelItems.value.find((m) => m.name === selectedModel.value)
+  if (!opt) return
+  savingModel.value = true
+  try {
+    await ml.patchSettings({
+      embedder_model_name: opt.name,
+      embedder_model_version: opt.version,
+    })
+    toast({ text: 'Embedding model saved — now Re-embed library, then Retrain heads', type: 'success' })
+  } catch (e) {
+    toast({ text: `Could not save model: ${e.message}`, type: 'error' })
+  } finally {
+    savingModel.value = false
+  }
+}
+
+async function onReembed() {
+  reembedding.value = true
+  try {
+    await store.backfill('embed')
+    await store.backfill('siglip')
+    toast({ text: 'Queued whole-image + concept re-embed — run the agent, then Retrain heads', type: 'success' })
+    await refreshQueue()
+  } catch (e) {
+    toast({ text: `Could not queue re-embed: ${e.message}`, type: 'error' })
+  } finally {
+    reembedding.value = false
+  }
+}
+
 async function onSaveAuto() {
   savingAuto.value = true
   try {
@@ -239,6 +333,37 @@ async function onBackfillSiglip() {
     backfillingSiglip.value = false
   }
 }
+
+async function onRetryErrors() {
+  retrying.value = true
+  try {
+    const { requeued, pruned, defects_kept: kept } = await store.retryErrors()
+    const extras = []
+    if (pruned) extras.push(`${pruned} stale duplicate${pruned === 1 ? '' : 's'} pruned`)
+    if (kept) extras.push(`${kept} known-bad file${kept === 1 ? '' : 's'} kept for recovery`)
+    const extra = extras.length ? ` (${extras.join(', ')})` : ''
+    toast({ text: `Requeued ${requeued} errored job${requeued === 1 ? '' : 's'}${extra} — run the agent to process them`, type: 'success' })
+    await refreshQueue()
+  } catch (e) {
+    toast({ text: `Could not retry errored jobs: ${e.message}`, type: 'error' })
+  } finally {
+    retrying.value = false
+  }
+}
+
+async function onReprocess() {
+  if (!window.confirm('Re-process the ENTIRE library (re-detect + re-crop every image)? This is heavy and runs on the GPU agent.')) return
+  reprocessing.value = true
+  try {
+    await store.reprocess('ccip')
+    toast({ text: 'Library queued for re-processing — run the agent to drain it', type: 'success' })
+    await refreshQueue()
+  } catch (e) {
+    toast({ text: `Could not start re-process: ${e.message}`, type: 'error' })
+  } finally {
+    reprocessing.value = false
+  }
+}
 </script>
 
 <style scoped>

frontend/src/components/settings/GpuTriageCard.vue

@@ -0,0 +1,189 @@
+<template>
+  <MaintenanceTile
+    icon="mdi-file-alert"
+    title="Failed processing"
+    blurb="Triage originals that failed GPU processing — probe the files, flag defects, recover them."
+  >
+    <p class="fc-muted text-body-2 mb-3">
+      A job that keeps failing parks as an error with its reason. A background
+      probe then checks the FILE itself (checksum + decode) and splits the
+      errors: <b>defective files</b> (truncated/corrupt originals — listed below
+      for recovery) vs <b>file&nbsp;OK</b> (the failure was operational; requeue
+      those with <i>Retry errored jobs</i> on the GPU agent card).
+    </p>
+
+    <div v-if="loading" class="fc-muted text-body-2">Loading…</div>
+    <template v-else-if="overview">
+      <div class="fc-queue mb-3">
+        <div class="fc-q"><div class="fc-q__n">{{ overview.total }}</div><div class="fc-q__l">errored</div></div>
+        <div class="fc-q"><div class="fc-q__n" :class="overview.triage.defect ? 'fc-weak' : ''">{{ overview.triage.defect }}</div><div class="fc-q__l">defective</div></div>
+        <div class="fc-q"><div class="fc-q__n fc-good">{{ overview.triage.file_ok }}</div><div class="fc-q__l">file ok</div></div>
+        <div class="fc-q"><div class="fc-q__n">{{ overview.triage.unclassified }}</div><div class="fc-q__l">unprobed</div></div>
+      </div>
+
+      <p v-if="classSummary" class="fc-muted text-caption mb-3">
+        Reasons: {{ classSummary }}
+      </p>
+
+      <div class="d-flex mb-4" style="gap:8px">
+        <v-btn
+          size="small" color="accent" variant="tonal" rounded="pill"
+          prepend-icon="mdi-magnify-scan" :loading="probing"
+          :disabled="!overview.triage.unclassified" @click="onProbe"
+        >Probe unclassified now</v-btn>
+        <v-btn
+          size="small" variant="text" rounded="pill"
+          prepend-icon="mdi-refresh" @click="refresh"
+        >Refresh</v-btn>
+      </div>
+
+      <template v-if="defects.length">
+        <div class="fc-section-h mb-2">Defective originals</div>
+        <div v-for="it in defects" :key="it.job_id" class="fc-defect mb-2">
+          <a :href="it.image_url" target="_blank" rel="noopener" class="fc-defect__thumb">
+            <img v-if="it.thumbnail_url" :src="it.thumbnail_url" alt="">
+            <v-icon v-else icon="mdi-file-question" size="28" />
+          </a>
+          <div class="fc-defect__meta">
+            <div class="text-body-2">
+              image <b>{{ it.image_id }}</b> · {{ it.task }} ·
+              <span class="fc-weak">{{ it.integrity_status }}</span>
+            </div>
+            <div class="fc-muted text-caption fc-defect__err" :title="it.error || ''">
+              {{ it.error || 'no stored reason' }}
+            </div>
+          </div>
+          <v-btn
+            v-if="recovered[it.image_id] !== 'no_source'"
+            size="small" color="accent" variant="tonal" rounded="pill"
+            prepend-icon="mdi-cloud-download" :loading="recovering === it.image_id"
+            @click="onRecover(it)"
+          >Recover</v-btn>
+          <span v-else class="fc-muted text-caption">
+            no pollable source — replace the file manually
+          </span>
+        </div>
+        <p class="fc-muted text-caption mt-2 mb-0">
+          Recover deletes the bad copy (and its record) and re-checks its
+          subscription source, so a fresh download re-imports it and re-enters
+          processing. Files without a pollable source need manual replacement.
+        </p>
+      </template>
+      <p v-else-if="!overview.total" class="fc-muted text-body-2 mb-0">
+        No failed jobs — the pipeline is clean.
+      </p>
+      <p v-else-if="!overview.triage.unclassified" class="fc-muted text-body-2 mb-0">
+        No defective files — every probed failure was operational
+        (file&nbsp;OK). Requeue them from the GPU agent card.
+      </p>
+    </template>
+  </MaintenanceTile>
+</template>
+
+<script setup>
+import { computed, onMounted, ref } from 'vue'
+
+import { toast } from '../../utils/toast.js'
+import MaintenanceTile from '../common/MaintenanceTile.vue'
+import { useGpuStore } from '../../stores/gpu.js'
+
+const store = useGpuStore()
+const loading = ref(true)
+const overview = ref(null)
+const probing = ref(false)
+const recovering = ref(null)
+// image_id -> 'no_source' for rows recovery already declined; keeps the
+// verdict visible instead of a button that fails the same way again.
+const recovered = ref({})
+
+const defects = computed(() =>
+  (overview.value?.items || []).filter((i) => i.triage_status === 'defect'))
+
+const classSummary = computed(() => {
+  const bc = overview.value?.by_class || {}
+  return Object.entries(bc)
+    .sort((a, b) => b[1] - a[1])
+    .map(([k, n]) => `${k.replaceAll('_', ' ')} ${n}`)
+    .join(' · ')
+})
+
+onMounted(refresh)
+
+async function refresh() {
+  loading.value = true
+  try {
+    overview.value = await store.errors()
+  } catch (e) {
+    toast({ text: `Could not load failed jobs: ${e.message}`, type: 'error' })
+  } finally {
+    loading.value = false
+  }
+}
+
+async function onProbe() {
+  probing.value = true
+  try {
+    await store.triageErrors()
+    toast({ text: 'Probe queued — verdicts appear here as files are checked (large videos take a while)', type: 'success' })
+  } catch (e) {
+    toast({ text: `Could not start the probe: ${e.message}`, type: 'error' })
+  } finally {
+    probing.value = false
+  }
+}
+
+async function onRecover(it) {
+  recovering.value = it.image_id
+  try {
+    const res = await store.recoverImage(it.image_id)
+    if (res.status === 'refetch_queued') {
+      toast({ text: `Deleted the bad copy and queued a re-check of source #${res.source_id} — it re-imports on the next fetch`, type: 'success' })
+      await refresh()
+    } else if (res.status === 'no_source') {
+      recovered.value = { ...recovered.value, [it.image_id]: 'no_source' }
+      toast({ text: 'No enabled subscription source covers this file — replace it manually', type: 'warning' })
+    } else {
+      toast({ text: 'Image record no longer exists — refreshing', type: 'warning' })
+      await refresh()
+    }
+  } catch (e) {
+    toast({ text: `Recovery failed: ${e.message}`, type: 'error' })
+  } finally {
+    recovering.value = null
+  }
+}
+</script>
+
+<style scoped>
+.fc-muted { color: rgb(var(--v-theme-on-surface-variant)); }
+.fc-section-h {
+  font-size: 13px; font-weight: 700; letter-spacing: 0.03em;
+  text-transform: uppercase; color: rgb(var(--v-theme-on-surface));
+}
+.fc-queue { display: flex; gap: 24px; }
+.fc-q__n {
+  font-size: 20px; font-weight: 700; line-height: 1.1;
+  font-family: 'JetBrains Mono', monospace;
+}
+.fc-q__l {
+  font-size: 11px; text-transform: uppercase; letter-spacing: 0.04em;
+  color: rgb(var(--v-theme-on-surface-variant));
+}
+.fc-good { color: rgb(var(--v-theme-success)); }
+.fc-weak { color: rgb(var(--v-theme-error)); }
+.fc-defect {
+  display: flex; align-items: center; gap: 12px;
+  background: rgb(var(--v-theme-surface-light)); border-radius: 8px;
+  padding: 6px 10px;
+}
+.fc-defect__thumb {
+  flex: 0 0 44px; width: 44px; height: 44px; border-radius: 6px;
+  overflow: hidden; display: flex; align-items: center; justify-content: center;
+  background: rgba(0, 0, 0, 0.25);
+}
+.fc-defect__thumb img { width: 100%; height: 100%; object-fit: cover; }
+.fc-defect__meta { flex: 1; min-width: 0; }
+.fc-defect__err {
+  overflow: hidden; text-overflow: ellipsis; white-space: nowrap;
+}
+</style>

frontend/src/components/settings/HeadsCard.vue

@@ -2,8 +2,8 @@
   <MaintenanceTile
     icon="mdi-brain"
     title="Concept heads (the learning suggester)"
-    blurb="Train the per-concept heads that turn your tags into suggestions — they replace Camie and sharpen every time you accept or reject."
+    blurb="Train the per-concept heads that turn your tags into suggestions — they learn from your library and sharpen every time you accept or reject."
-    :open="headCount > 0 || running"
+    :open="running"
   >
     <p class="fc-muted text-body-2 mb-3">
       A <strong>head</strong> is a tiny classifier trained on the SigLIP

frontend/src/components/settings/ImportFiltersForm.vue

@@ -90,10 +90,14 @@
 </template>
 
 <script setup>
-import { reactive, watch } from 'vue'
+import { onMounted, reactive, watch } from 'vue'
 import { useImportStore } from '../../stores/import.js'
 
 const store = useImportStore()
+// Self-sufficient since the Import tab dissolved (2026-07-02): this form now
+// lives in Maintenance → Ingestion & filters and loads its own settings
+// instead of relying on the old tab's mount hook.
+onMounted(() => { if (!store.settings) store.loadSettings() })
 // Labelled stops so the less-initiated get the gist without knowing what a
 // Hamming distance is. 0 = byte-for-byte only; 10 = the shipped default.
 const PHASH_TICKS = { 0: 'Exact', 4: 'Strict', 10: 'Default', 16: 'Loose' }

frontend/src/components/settings/ImportTaskList.vue

@@ -1,254 +0,0 @@
-<template>
-  <v-card>
-    <CardHeading title="Recent import tasks">
-      <v-spacer />
-      <v-select
-        v-model="statusFilter" :items="statusOptions" density="compact"
-        hide-details style="max-width: 180px;" @update:model-value="onFilterChange"
-      />
-      <v-btn variant="text" rounded="pill" size="small" @click="onRefresh">
-        <v-icon start>mdi-refresh</v-icon>
-        Refresh
-      </v-btn>
-      <v-btn
-        variant="text" rounded="pill" size="small" color="warning"
-        :disabled="!hasFailed" @click="onRetryFailed"
-      >
-        Retry failed
-      </v-btn>
-      <v-btn
-        variant="text" rounded="pill" size="small" color="warning"
-        :disabled="!hasStuck" @click="onClearStuckOpen"
-      >
-        Clear stuck…
-      </v-btn>
-      <v-btn
-        variant="text" rounded="pill" size="small" color="error"
-        @click="onClearOpen"
-      >
-        Clear completed…
-      </v-btn>
-    </CardHeading>
-    <v-data-table-virtual
-      :headers="headers" :items="store.tasks" :loading="store.tasksLoading"
-      height="480" density="compact" fixed-header no-data-text="No tasks yet — trigger a scan above."
-    >
-      <template #item.status="{ item }">
-        <v-chip :color="statusColor(item.status)" size="small" variant="tonal">
-          {{ item.status }}
-        </v-chip>
-      </template>
-      <template #item.source_path="{ item }">
-        <span :title="item.source_path">{{ shorten(item.source_path) }}</span>
-      </template>
-      <template #item.size_bytes="{ item }">{{ formatBytes(item.size_bytes) }}</template>
-      <template #item.created_at="{ item }">{{ formatDate(item.created_at) }}</template>
-      <template #item.error="{ item }">
-        <button
-          v-if="item.error" type="button" class="fc-err-link text-caption"
-          @click="openError(`Task ${item.id} failed`, item.error)"
-          title="Click for full error"
-        >{{ shorten(item.error, 60) }}</button>
-      </template>
-      <template #item.actions="{ item }">
-        <v-btn
-          v-if="item.status === 'failed'"
-          icon size="x-small" variant="text"
-          :loading="refetching === item.id"
-          @click="onRefetch(item)"
-        >
-          <v-icon size="small">mdi-cloud-refresh</v-icon>
-          <v-tooltip activator="parent" location="top">
-            Re-fetch original (re-download from source)
-          </v-tooltip>
-        </v-btn>
-      </template>
-    </v-data-table-virtual>
-    <div v-if="store.hasMore" class="d-flex justify-center py-3">
-      <v-btn variant="text" size="small" @click="onLoadMore">Load more</v-btn>
-    </div>
-
-    <v-dialog v-model="clearDialog" max-width="400">
-      <v-card>
-        <v-card-title>Clear completed tasks</v-card-title>
-        <v-card-text>
-          <v-select
-            v-model="clearAgeDays" label="Older than"
-            :items="[
-              { title: 'All finished', value: 0 },
-              { title: '1 day',  value: 1 },
-              { title: '7 days', value: 7 },
-              { title: '30 days', value: 30 }
-            ]"
-          />
-        </v-card-text>
-        <v-card-actions>
-          <v-spacer />
-          <v-btn @click="clearDialog = false">Cancel</v-btn>
-          <v-btn color="error" rounded="pill" @click="onClearConfirm">Clear</v-btn>
-        </v-card-actions>
-      </v-card>
-    </v-dialog>
-
-    <v-dialog v-model="clearStuckDialog" max-width="480">
-      <v-card>
-        <v-card-title>Clear stuck tasks</v-card-title>
-        <v-card-text>
-          <v-alert type="warning" variant="tonal" density="compact" class="mb-3">
-            Force every <strong>pending / queued / processing</strong> task to
-            <strong>failed</strong> and finalize any active batch that
-            has no remaining work. Use this when the automatic recovery
-            sweep keeps re-queueing the same row (e.g., corrupt file in
-            an autoretry loop, or worker model missing).
-          </v-alert>
-          <p class="text-body-2">
-            Tasks remain in the database with status=<code>failed</code>;
-            click <em>Retry failed</em> once the underlying cause is
-            resolved to re-queue them.
-          </p>
-        </v-card-text>
-        <v-card-actions>
-          <v-spacer />
-          <v-btn @click="clearStuckDialog = false">Cancel</v-btn>
-          <v-btn color="warning" rounded="pill" @click="onClearStuckConfirm">Clear stuck</v-btn>
-        </v-card-actions>
-      </v-card>
-    </v-dialog>
-
-    <ErrorDetailModal
-      v-model="showErrorModal"
-      :title="errorModalTitle"
-      :message="errorModalMessage"
-    />
-  </v-card>
-</template>
-
-<script setup>
-import { toast } from '../../utils/toast.js'
-import { computed, ref } from 'vue'
-import { useImportStore } from '../../stores/import.js'
-import ErrorDetailModal from '../common/ErrorDetailModal.vue'
-import CardHeading from '../common/CardHeading.vue'
-
-const store = useImportStore()
-const statusFilter = ref(null)
-const clearDialog = ref(false)
-// Click-to-open modal for full error text (operator-flagged 2026-05-26
-// — the prior :title="..." tooltip cramped multi-line SQLAlchemy
-// tracebacks into an unusable popup with no copy-paste affordance).
-const showErrorModal = ref(false)
-const errorModalTitle = ref('')
-const errorModalMessage = ref('')
-
-function openError(title, message) {
-  errorModalTitle.value = title
-  errorModalMessage.value = message || ''
-  showErrorModal.value = true
-}
-const clearAgeDays = ref(7)
-const clearStuckDialog = ref(false)
-
-const statusOptions = [
-  { title: 'All', value: null },
-  { title: 'Pending', value: 'pending' },
-  { title: 'Queued', value: 'queued' },
-  { title: 'Processing', value: 'processing' },
-  { title: 'Complete', value: 'complete' },
-  { title: 'Skipped', value: 'skipped' },
-  { title: 'Failed', value: 'failed' }
-]
-
-const headers = [
-  { title: 'Status', key: 'status', sortable: false, width: 120 },
-  { title: 'Source', key: 'source_path', sortable: false },
-  { title: 'Size', key: 'size_bytes', sortable: false, width: 90 },
-  { title: 'Created', key: 'created_at', sortable: false, width: 150 },
-  { title: 'Note', key: 'error', sortable: false },
-  { title: '', key: 'actions', sortable: false, width: 56 }
-]
-
-const refetching = ref(null)
-const _REFETCH_MSG = {
-  refetch_queued: { text: 'Re-fetch queued — re-downloading from source', type: 'success' },
-  no_source: { text: 'No re-fetchable source (filesystem import — replace the file manually)', type: 'info' },
-  already_refetched: { text: 'Already re-fetched once', type: 'info' },
-}
-async function onRefetch(item) {
-  refetching.value = item.id
-  try {
-    const res = await store.refetchTask(item.id)
-    const msg = _REFETCH_MSG[res.status] || { text: `Re-fetch: ${res.status}`, type: 'info' }
-    toast(msg)
-  } catch (e) {
-    toast({ text: `Re-fetch failed: ${e.message}`, type: 'error' })
-  } finally {
-    refetching.value = null
-  }
-}
-
-const hasFailed = computed(() => store.tasks.some(t => t.status === 'failed'))
-const hasStuck = computed(() => store.tasks.some(
-  t => t.status === 'pending' || t.status === 'queued' || t.status === 'processing'
-))
-
-function statusColor(s) {
-  return {
-    complete: 'success',
-    skipped: 'warning',
-    failed: 'error',
-    processing: 'accent',
-    queued: 'info',
-    pending: 'info'
-  }[s] || 'default'
-}
-function shorten(s, max = 90) {
-  if (!s) return ''
-  if (s.length <= max) return s
-  const head = Math.floor((max - 3) * 0.6)
-  const tail = max - 3 - head
-  return s.slice(0, head) + '...' + s.slice(-tail)
-}
-function formatBytes(b) {
-  if (!b) return ''
-  const units = ['B', 'KiB', 'MiB', 'GiB']
-  let i = 0; let v = b
-  while (v >= 1024 && i < units.length - 1) { v /= 1024; i++ }
-  return `${v.toFixed(i === 0 ? 0 : 1)} ${units[i]}`
-}
-function formatDate(s) {
-  try { return new Date(s).toLocaleString() } catch { return s }
-}
-
-async function onRefresh() { await store.loadTasks(true) }
-function onFilterChange() { store.setStatusFilter(statusFilter.value); store.loadTasks(true) }
-async function onLoadMore() { await store.loadTasks(false) }
-async function onRetryFailed() { await store.retryFailed() }
-function onClearOpen() { clearDialog.value = true }
-async function onClearConfirm() {
-  await store.clearCompleted(clearAgeDays.value)
-  clearDialog.value = false
-}
-function onClearStuckOpen() { clearStuckDialog.value = true }
-async function onClearStuckConfirm() {
-  await store.clearStuck()
-  clearStuckDialog.value = false
-}
-</script>
-
-<style scoped>
-.fc-err-link {
-  /* Truncated error preview as a clickable button — opens
-     ErrorDetailModal with the full text. Inherits the row's font
-     sizing so it doesn't visually drift from the prior tooltip-bearing
-     span. */
-  color: rgb(var(--v-theme-error, 220 80 80));
-  background: transparent;
-  border: 0;
-  padding: 0;
-  font: inherit;
-  text-align: left;
-  text-decoration: underline dotted;
-  cursor: pointer;
-}
-.fc-err-link:hover { text-decoration: underline; }
-</style>

frontend/src/components/settings/ImportTriggerPanel.vue

@@ -1,97 +0,0 @@
-<template>
-  <v-card>
-    <v-card-title>Trigger scan</v-card-title>
-    <v-card-text>
-      <div v-if="store.activeBatch" class="d-flex align-center mb-3" style="gap: 12px;">
-        <v-progress-circular
-          indeterminate color="accent" size="20"
-        />
-        <span>
-          {{ store.activeBatch.scan_mode === 'deep' ? 'Deep scanning' : 'Scanning' }}
-          {{ store.activeBatch.source_path || '/import' }} —
-          imported {{ store.activeBatch.imported }},
-          <template v-if="store.activeBatch.scan_mode === 'deep'">
-            refreshed {{ store.activeBatch.refreshed || 0 }},
-          </template>
-          skipped {{ store.activeBatch.skipped }},
-          failed {{ store.activeBatch.failed }} /
-          {{ store.activeBatch.total_files }} files
-        </span>
-        <v-spacer />
-        <v-btn
-          variant="text" rounded="pill" size="small" color="warning"
-          :loading="clearing" @click="onClearStuck"
-        >
-          Clear stuck
-        </v-btn>
-      </div>
-
-      <p class="text-body-2 mb-3">
-        <span v-if="!store.activeBatch">
-          <strong>Quick scan</strong> walks <code>/import</code> and enqueues
-          new files only.
-          <strong>Deep scan</strong> additionally re-walks already-imported
-          files so updated sidecar metadata (post title/date/attribution) and
-          previously-NULL phashes / artist links get refreshed. Use after
-          bulk-downloading fresh sidecars for existing content. Both modes
-          route non-media + sidecar pairs through PostAttachment capture.
-        </span>
-        <span v-else>
-          An active batch is in progress. Wait for it to finish, or click
-          <em>Clear stuck</em> above if it has been wedged with no
-          measurable progress.
-        </span>
-      </p>
-
-      <div class="d-flex flex-wrap" style="gap: 12px;">
-        <v-btn
-          color="primary" rounded="pill"
-          :disabled="!!store.activeBatch"
-          :loading="busy === 'quick'"
-          @click="trigger('quick')"
-        >
-          <v-icon start>mdi-magnify-scan</v-icon>
-          Quick scan
-        </v-btn>
-        <v-btn
-          color="secondary" rounded="pill" variant="tonal"
-          :disabled="!!store.activeBatch"
-          :loading="busy === 'deep'"
-          @click="trigger('deep')"
-        >
-          <v-icon start>mdi-magnify-plus-outline</v-icon>
-          Deep scan
-        </v-btn>
-      </div>
-
-      <v-alert v-if="store.triggerError" type="error" variant="tonal" class="mt-3" closable>
-        {{ store.triggerError }}
-      </v-alert>
-    </v-card-text>
-  </v-card>
-</template>
-
-<script setup>
-import { ref } from 'vue'
-import { useImportStore } from '../../stores/import.js'
-
-const store = useImportStore()
-const busy = ref(null)
-const clearing = ref(false)
-
-async function trigger(mode) {
-  busy.value = mode
-  try { await store.triggerScan(mode) } catch {} finally { busy.value = null }
-}
-
-async function onClearStuck() {
-  clearing.value = true
-  try {
-    await store.clearStuck()
-  } catch {
-    // store surfaces error via triggerError if needed
-  } finally {
-    clearing.value = false
-  }
-}
-</script>

frontend/src/components/settings/MLBackfillCard.vue

@@ -1,15 +1,33 @@
 <template>
   <MaintenanceTile
     icon="mdi-refresh"
-    title="ML backfill"
+    title="CPU embedding backfill"
-    blurb="Re-run tagging + embeddings on images missing them."
+    blurb="Whole-image embeddings without a GPU agent — the built-in fallback."
     :open="busy"
   >
       <p class="text-body-2 mb-3">
-        Re-run Camie + SigLIP on images missing predictions or embeddings
+        Computes the whole-image SigLIP embedding for anything missing one —
-        for the current model versions. Safe to re-run.
+        images directly, videos by sampling frames (the same approach as the
+        GPU agent). Runs on the ml-worker's CPU, so search, similarity and
+        head suggestions work <strong>without</strong> a GPU agent; new imports
+        are embedded this way automatically. Detection, cropping and character
+        (CCIP) embeddings are GPU-agent-only. Safe to re-run. To re-embed under
+        a NEW model, use the GPU agent's "Re-embed library" instead.
       </p>
-      <v-btn color="primary" rounded="pill" :loading="busy" @click="run">
+      <v-switch
+        v-model="enabled" color="accent" hide-details density="compact"
+        :loading="saving" label="CPU embedding enabled"
+        class="mb-1" @update:model-value="onToggle"
+      />
+      <p class="fc-muted text-caption mb-3">
+        Turn OFF if you run the GPU agent and removed the ml-worker container —
+        imports then stop queueing CPU embed work nothing will consume (the
+        daily GPU embed backfill covers those images instead).
+      </p>
+      <v-btn
+        color="primary" rounded="pill" :loading="busy" :disabled="!enabled"
+        @click="run"
+      >
         <v-icon start>mdi-refresh</v-icon> Run backfill now
       </v-btn>
       <span v-if="done" class="ml-3 text-caption">Enqueued.</span>
@@ -19,13 +37,40 @@
 
 <script setup>
 import { toast } from '../../utils/toast.js'
-import { ref } from 'vue'
+import { onMounted, ref } from 'vue'
 import { useMLStore } from '../../stores/ml.js'
 import MaintenanceTile from '../common/MaintenanceTile.vue'
 import QueueStatusBar from './QueueStatusBar.vue'
 const store = useMLStore()
 const busy = ref(false)
 const done = ref(false)
+const enabled = ref(true)
+const saving = ref(false)
+onMounted(async () => {
+  try {
+    await store.loadSettings()
+    if (store.settings?.cpu_embed_enabled != null) {
+      enabled.value = store.settings.cpu_embed_enabled
+    }
+  } catch { /* non-fatal */ }
+})
+async function onToggle() {
+  saving.value = true
+  try {
+    await store.patchSettings({ cpu_embed_enabled: enabled.value })
+    toast({
+      text: enabled.value
+        ? 'CPU embedding on — imports queue embeds for the ml-worker'
+        : 'CPU embedding off — the GPU embed backfill owns whole-image embeds',
+      type: 'success',
+    })
+  } catch (e) {
+    toast({ text: `Could not save: ${e.message}`, type: 'error' })
+    enabled.value = !enabled.value
+  } finally {
+    saving.value = false
+  }
+}
 async function run() {
   busy.value = true
   try { await store.triggerBackfill(); done.value = true }
@@ -33,3 +78,7 @@ async function run() {
   finally { busy.value = false }
 }
 </script>
+
+<style scoped>
+.fc-muted { color: rgb(var(--v-theme-on-surface-variant)); }
+</style>

frontend/src/components/settings/MLThresholdSliders.vue

@@ -1,70 +1,30 @@
 <template>
   <MaintenanceTile
-    icon="mdi-tune"
+    icon="mdi-filmstrip"
-    title="Suggestion thresholds"
+    title="Video embedding"
-    blurb="Confidence cutoffs that gate auto-suggested tags + video sampling."
+    blurb="How videos are sampled into frames before embedding."
   >
     <div v-if="store.settings">
-      <v-row v-for="f in fields" :key="f.key">
-        <v-col cols="12">
-          <v-slider
-            v-model="local[f.key]" :label="f.label"
-            :min="f.floorMin ? local.tagger_store_floor : 0" max="1" step="0.05"
-            thumb-label hide-details
-            color="accent" @end="save"
-          />
-        </v-col>
-      </v-row>
-
-      <v-divider class="my-4" />
-
-      <v-row>
-        <v-col cols="12">
-          <v-slider
-            v-model="local.tagger_store_floor" label="Tagger store floor"
-            min="0" max="1" step="0.05" thumb-label hide-details
-            color="accent" @end="save"
-          />
-          <div class="text-caption fc-muted mt-1">
-            Tagger predictions below this confidence aren't stored — raising it
-            keeps the image library lean. Suggestions can't be shown below the
-            floor; lower-confidence tags you actually want still surface through
-            the learned centroid path.
-          </div>
-        </v-col>
-      </v-row>
-
-      <v-divider class="my-4" />
-
-      <div class="text-subtitle-2 mb-1">Video tagging</div>
       <div class="text-caption fc-muted mb-3">
-        Videos are tagged by sampling frames at a fixed cadence. A tag is kept
+        Videos are embedded by sampling frames at a fixed cadence and mean-pooling
-        only if it shows up in enough frames (≈ that many × the interval in
+        their SigLIP embeddings. The interval sets the cadence; the cap bounds how
-        seconds of screen time), which filters one-frame noise without losing
+        many frames a long video samples.
-        tags that only appear in part of a longer video.
       </div>
       <v-row>
-        <v-col cols="12" sm="4">
+        <v-col cols="12" sm="6">
           <v-text-field
             v-model.number="local.video_frame_interval_seconds"
             label="Frame interval (s)" type="number" min="0.5" step="0.5"
             density="comfortable" hide-details @change="save"
           />
         </v-col>
-        <v-col cols="12" sm="4">
+        <v-col cols="12" sm="6">
           <v-text-field
             v-model.number="local.video_max_frames"
             label="Max frames" type="number" min="1" step="1"
             density="comfortable" hide-details @change="save"
           />
         </v-col>
-        <v-col cols="12" sm="4">
-          <v-text-field
-            v-model.number="local.video_min_tag_frames"
-            label="Min frames per tag" type="number" min="1" step="1"
-            density="comfortable" hide-details @change="save"
-          />
-        </v-col>
       </v-row>
     </div>
     <div v-else><v-skeleton-loader type="paragraph" /></div>
@@ -78,32 +38,14 @@ import { useMLStore } from '../../stores/ml.js'
 import MaintenanceTile from '../common/MaintenanceTile.vue'
 
 const store = useMLStore()
-// 'artist' (FC-2d-vii-c) and 'copyright' (2026-06-01) retired as
-// suggestion categories; their threshold rows are gone.
-// floorMin: the per-category suggestion thresholds can't drop below the
-// tagger store floor (nothing below the floor is stored to surface).
-const fields = [
-  { key: 'suggestion_threshold_character', label: 'Character', floorMin: true },
-  { key: 'suggestion_threshold_general', label: 'General', floorMin: true },
-  { key: 'centroid_similarity_threshold', label: 'Centroid similarity' }
-]
 const local = reactive({})
 watch(() => store.settings, (s) => { if (s) Object.assign(local, s) }, { immediate: true })
 
 async function save() {
-  // Mirror the server invariant: keep the category thresholds at or above the
+  const patch = {
-  // store floor so a raised floor doesn't leave a threshold stranded below it.
+    video_frame_interval_seconds: local.video_frame_interval_seconds,
-  const floor = local.tagger_store_floor
+    video_max_frames: local.video_max_frames
-  local.suggestion_threshold_character = Math.max(local.suggestion_threshold_character, floor)
+  }
-  local.suggestion_threshold_general = Math.max(local.suggestion_threshold_general, floor)
-  // Mirror the server invariant: a tag can't require more frames than are sampled.
-  local.video_min_tag_frames = Math.min(local.video_min_tag_frames, local.video_max_frames)
-  const patch = {}
-  for (const f of fields) patch[f.key] = local[f.key]
-  patch.tagger_store_floor = local.tagger_store_floor
-  patch.video_frame_interval_seconds = local.video_frame_interval_seconds
-  patch.video_max_frames = local.video_max_frames
-  patch.video_min_tag_frames = local.video_min_tag_frames
   try { await store.patchSettings(patch) }
   catch (e) { toast({ text: e.message, type: 'error' }) }
 }

frontend/src/components/settings/MaintenancePanel.vue

@@ -1,36 +1,56 @@
 <template>
   <div class="fc-maint">
     <p class="fc-muted text-body-2 mb-5">
-      One-off backfills, tagging config and storage tools. The ML backfill and
+      Processing, tagging and storage tools, grouped by system. Heads train
-      centroid recompute also run nightly; the allowlist auto-applies accepted
+      nightly and auto-apply earned tags. Click a tile to open it.
-      tags. Click a tile to open it.
     </p>
 
     <section class="fc-section">
-      <h3 class="fc-section__title">Backfills &amp; reprocessing</h3>
+      <h3 class="fc-section__title">Ingestion &amp; filters</h3>
-      <p class="fc-section__hint">Re-run tagging, thumbnails, extraction and DB upkeep.</p>
+      <p class="fc-section__hint">
-      <div class="fc-tile-grid">
+        What gets imported — dedup sensitivity, size/transparency/solid-color
+        filters. Applies to downloads and folder imports alike.
+      </p>
+      <div class="fc-tile-stack">
+        <ImportFiltersForm />
+      </div>
+    </section>
+
+    <section class="fc-section">
+      <h3 class="fc-section__title">GPU agent &amp; embeddings</h3>
+      <p class="fc-section__hint">
+        The desktop agent that does the heavy lifting, its failure triage, and
+        the CPU fallback.
+      </p>
+      <div class="fc-tile-stack">
+        <GpuAgentCard />
+        <GpuTriageCard />
         <MLBackfillCard />
-        <CentroidRecomputeCard />
-        <ThumbnailBackfillCard />
-        <ArchiveReextractCard />
-        <MissingFileRepairCard />
-        <DbMaintenanceCard />
       </div>
     </section>
 
     <section class="fc-section">
       <h3 class="fc-section__title">Tagging</h3>
       <p class="fc-section__hint">
-        Suggestion thresholds, the auto-apply allowlist and tag aliases.
+        Suggestion thresholds, trained heads and tag aliases.
       </p>
       <div class="fc-tile-stack">
         <MLThresholdSliders />
         <HeadsCard />
-        <GpuAgentCard />
-        <AllowlistTable />
         <AliasTable />
-        <TagEvalCard />
+      </div>
+    </section>
+
+    <section class="fc-section">
+      <h3 class="fc-section__title">Library health</h3>
+      <p class="fc-section__hint">
+        Self-healing and repair: missing files, thumbnails, database upkeep.
+      </p>
+      <div class="fc-tile-grid">
+        <MissingFileRepairCard />
+        <ThumbnailBackfillCard />
+        <DbMaintenanceCard />
+        <ArchiveReextractCard />
       </div>
     </section>
 
@@ -47,18 +67,17 @@
 <script setup>
 import { onMounted, onUnmounted } from 'vue'
 
+import ImportFiltersForm from './ImportFiltersForm.vue'
 import MLBackfillCard from './MLBackfillCard.vue'
-import CentroidRecomputeCard from './CentroidRecomputeCard.vue'
 import ThumbnailBackfillCard from './ThumbnailBackfillCard.vue'
 import ArchiveReextractCard from './ArchiveReextractCard.vue'
 import MissingFileRepairCard from './MissingFileRepairCard.vue'
+import GpuTriageCard from './GpuTriageCard.vue'
 import DbMaintenanceCard from './DbMaintenanceCard.vue'
 import MLThresholdSliders from './MLThresholdSliders.vue'
 import HeadsCard from './HeadsCard.vue'
 import GpuAgentCard from './GpuAgentCard.vue'
-import AllowlistTable from './AllowlistTable.vue'
 import AliasTable from './AliasTable.vue'
-import TagEvalCard from './TagEvalCard.vue'
 import BackupCard from './BackupCard.vue'
 import { useSystemActivityStore } from '../../stores/systemActivity.js'
 

frontend/src/components/settings/SystemActivityTab.vue

@@ -17,6 +17,12 @@
       </v-card-text>
     </v-card>
 
+    <!-- The non-Celery halves of the app (2026-07-02): the GPU agent does the
+         majority of processing and downloads feed the library — Activity is
+         the whole-app pulse, not just the worker queues. -->
+    <GpuActivityPanel @open-maintenance="$emit('open-maintenance')" />
+    <DownloadsActivityPanel />
+
     <!-- Recent failures pane -->
     <v-card class="mb-4">
       <CardHeading icon="mdi-alert-circle-outline" title="Recent failures (last 24h)">
@@ -167,6 +173,10 @@ import { formatRelative as fmtRelative } from '../../utils/date.js'
 import ErrorDetailModal from '../common/ErrorDetailModal.vue'
 import QueuesTable from './QueuesTable.vue'
 import CardHeading from '../common/CardHeading.vue'
+import GpuActivityPanel from './GpuActivityPanel.vue'
+import DownloadsActivityPanel from './DownloadsActivityPanel.vue'
+
+defineEmits(['open-maintenance'])
 
 // Click-to-open modal for full error text. Replaces the unusable
 // :title="..." tooltip (operator-flagged 2026-05-26: SQLAlchemy

frontend/src/components/settings/TagEvalCard.vue

@@ -1,303 +0,0 @@
-<template>
-  <MaintenanceTile
-    icon="mdi-flask-outline"
-    title="Tagging eval (heads vs centroid)"
-    blurb="Measure whether a trained head beats the old centroid on your own tags — and whether tagging more sharpens it."
-    :open="!!run"
-  >
-    <p class="fc-muted text-body-2 mb-3">
-      Reuses the SigLIP embeddings already stored on your images (no re-embed, no
-      GPU). For each concept it trains a logistic-regression <strong>head</strong>
-      on your positives + negatives and compares it to the old single
-      <strong>centroid</strong>, with cross-validated AP/F1 and a learning curve.
-      Runs as a background task; the result is saved and reloads here.
-    </p>
-
-    <v-textarea
-      v-model="conceptsText" label="Concepts (comma-separated)"
-      rows="2" auto-grow density="compact" hide-details class="mb-3"
-      :disabled="running"
-    />
-
-    <div class="d-flex mb-3" style="gap: 12px;">
-      <v-text-field
-        v-model.number="autoTopN" label="+ auto-add top-N concepts"
-        type="number" min="0" max="200" density="compact" hide-details
-        :disabled="running" style="max-width: 220px;"
-      />
-      <v-text-field
-        v-model.number="precisionTarget" label="Auto-apply precision target"
-        type="number" min="0.5" max="0.999" step="0.01" density="compact" hide-details
-        :disabled="running" style="max-width: 220px;"
-      />
-    </div>
-
-    <v-btn
-      v-if="!running"
-      color="accent" variant="flat" rounded="pill"
-      prepend-icon="mdi-play" :loading="busy" @click="onStart"
-    >Run eval</v-btn>
-
-    <div v-if="running" class="mt-3">
-      <v-progress-linear indeterminate color="accent" />
-      <div class="text-body-2 mt-2 fc-muted">Running… (started {{ startedAgo }})</div>
-    </div>
-
-    <v-alert
-      v-if="run && run.status === 'error'"
-      type="error" variant="tonal" density="compact" class="mt-3"
-    >Eval failed: {{ run.error }}</v-alert>
-
-    <div v-if="report" class="mt-4">
-      <div class="fc-muted text-caption mb-2">
-        Ran {{ formatTime(report.generated_at) }} ·
-        {{ report.concepts.length }} concept(s) ·
-        neg ratio {{ report.params.neg_ratio }}, {{ report.params.cv_folds }}-fold CV
-      </div>
-
-      <div v-for="c in report.concepts" :key="c.name" class="fc-cc">
-        <div class="fc-cc__head">
-          <span class="fc-cc__name">{{ c.name }}</span>
-          <span v-if="c.skipped" class="fc-muted text-caption">— skipped: {{ c.skipped }}</span>
-          <span v-else class="fc-muted text-caption">
-            {{ c.n_pos }} pos · {{ c.n_neg }} neg<span v-if="c.n_rejected"> ({{ c.n_rejected }} rejected)</span>
-          </span>
-        </div>
-
-        <template v-if="!c.skipped">
-          <table class="fc-metrics">
-            <thead>
-              <tr><th></th><th>AP</th><th>F1</th><th>Prec</th><th>Rec</th></tr>
-            </thead>
-            <tbody>
-              <tr>
-                <td class="fc-metrics__lbl">Head</td>
-                <td class="fc-num fc-win">{{ c.head.ap }}</td>
-                <td class="fc-num">{{ c.head.f1 }}</td>
-                <td class="fc-num">{{ c.head.precision }}</td>
-                <td class="fc-num">{{ c.head.recall }}</td>
-              </tr>
-              <tr>
-                <td class="fc-metrics__lbl fc-muted">Centroid</td>
-                <td class="fc-num fc-muted">{{ c.centroid.ap }}</td>
-                <td class="fc-num fc-muted">{{ c.centroid.f1 }}</td>
-                <td class="fc-num fc-muted">{{ c.centroid.precision }}</td>
-                <td class="fc-num fc-muted">{{ c.centroid.recall }}</td>
-              </tr>
-            </tbody>
-          </table>
-          <div class="text-caption mb-2" :class="apDelta(c) >= 0 ? 'fc-up' : 'fc-down'">
-            Δ AP {{ apDelta(c) >= 0 ? '+' : '' }}{{ apDelta(c).toFixed(3) }}
-            (head − centroid)
-          </div>
-
-          <div class="text-caption mb-2">
-            <span class="fc-muted">Auto-apply:</span>
-            <template v-if="c.head.auto_apply">
-              <span class="fc-up">ready</span> — at P≥{{ c.head.auto_apply.target }}
-              catches recall <strong>{{ c.head.auto_apply.recall }}</strong>
-              (thr {{ c.head.auto_apply.threshold }})
-            </template>
-            <span v-else class="fc-down">not reachable at P≥{{ report.params.precision_target }}</span>
-          </div>
-
-          <div v-if="c.curve && c.curve.length" class="fc-curve">
-            <span class="fc-muted text-caption">Learning curve (AP @ N positives):</span>
-            <span v-for="p in c.curve" :key="p.n_pos" class="fc-curve__pt">
-              {{ p.n_pos }}→<strong>{{ p.ap }}</strong>
-            </span>
-          </div>
-
-          <div v-if="c.examples" class="fc-ex">
-            <div
-              v-for="grp in [
-                { dir: 'suggest', items: c.examples.head_would_suggest,
-                  label: `Head would suggest — ✓ tag it, ✗ not ${c.name}` },
-                { dir: 'doubts', items: c.examples.head_doubts_positive,
-                  label: `Head doubts your tag — ✓ keep, ✗ remove (not ${c.name})` },
-              ]" :key="grp.dir" class="fc-ex__row"
-            >
-              <div class="fc-muted text-caption mb-1">{{ grp.label }}</div>
-              <div class="fc-ex__thumbs">
-                <div
-                  v-for="it in grp.items" :key="`${grp.dir}${it.id}`"
-                  class="fc-ex__item"
-                  :class="actedLabel(c, grp.dir, it) ? 'fc-ex__item--acted' : ''"
-                >
-                  <button
-                    type="button" class="fc-ex__thumb"
-                    :title="`#${it.id} — click to enlarge`" @click="modal.open(it.id)"
-                  >
-                    <img :src="it.thumbnail_url" loading="lazy" />
-                  </button>
-                  <div v-if="actedLabel(c, grp.dir, it)" class="fc-ex__badge">
-                    {{ actedLabel(c, grp.dir, it) }}
-                  </div>
-                  <div v-else class="fc-ex__acts">
-                    <button
-                      class="fc-act fc-act--yes" type="button"
-                      :title="`Yes — it is ${c.name}`" @click="act(c, it, grp.dir, 'yes')"
-                    ><v-icon size="15">mdi-check</v-icon></button>
-                    <button
-                      class="fc-act fc-act--no" type="button"
-                      :title="`No — not ${c.name}`" @click="act(c, it, grp.dir, 'no')"
-                    ><v-icon size="15">mdi-close</v-icon></button>
-                  </div>
-                </div>
-              </div>
-            </div>
-          </div>
-        </template>
-      </div>
-    </div>
-  </MaintenanceTile>
-</template>
-
-<script setup>
-import { toast } from '../../utils/toast.js'
-import { computed, onMounted, onUnmounted, ref } from 'vue'
-
-import MaintenanceTile from '../common/MaintenanceTile.vue'
-import { useTagEvalStore } from '../../stores/tagEval.js'
-import { useModalStore } from '../../stores/modal.js'
-
-const DEFAULT_CONCEPTS =
-  'glasses, cat, dog, horse, goblin, cum, lactation, fellatio, xray, stomach bulge'
-
-const store = useTagEvalStore()
-const modal = useModalStore()
-const run = ref(null)
-const conceptsText = ref(DEFAULT_CONCEPTS)
-const autoTopN = ref(0)
-const precisionTarget = ref(0.97)
-const busy = ref(false)
-let pollTimer = null
-
-const running = computed(() => run.value?.status === 'running')
-const report = computed(() => (run.value?.status === 'ready' ? run.value.report : null))
-const startedAgo = computed(() =>
-  run.value?.started_at ? formatTime(run.value.started_at) : '')
-
-// Rehydrate the persisted run on mount so the report survives navigation — the
-// task runs backend-side regardless; we just reconnect to its row.
-onMounted(async () => {
-  try {
-    const latest = await store.latest()
-    if (latest) {
-      run.value = await store.getRun(latest.id)
-      if (run.value.status === 'running') startPoll(latest.id)
-    }
-  } catch { /* non-fatal — card still works for a fresh run */ }
-})
-onUnmounted(stopPoll)
-
-function startPoll(id) {
-  stopPoll()
-  pollTimer = setInterval(async () => {
-    try {
-      run.value = await store.getRun(id)
-      if (run.value.status !== 'running') stopPoll()
-    } catch (e) {
-      stopPoll()
-      toast({ text: `Eval poll failed: ${e.message}`, type: 'error' })
-    }
-  }, 5000)
-}
-function stopPoll() {
-  if (pollTimer) { clearInterval(pollTimer); pollTimer = null }
-}
-
-async function onStart() {
-  busy.value = true
-  try {
-    const concepts = conceptsText.value.split(',').map(s => s.trim()).filter(Boolean)
-    const res = await store.start({
-      concepts,
-      auto_top_n: Number(autoTopN.value) || 0,
-      precision_target: Number(precisionTarget.value) || 0.97,
-    })
-    run.value = await store.getRun(res.run_id)
-    startPoll(res.run_id)
-  } catch (e) {
-    const msg = e.body?.running_id
-      ? 'An eval is already running.'
-      : e.message
-    toast({ text: `Could not start eval: ${msg}`, type: 'error' })
-  } finally {
-    busy.value = false
-  }
-}
-
-function apDelta(c) { return (c.head?.ap ?? 0) - (c.centroid?.ap ?? 0) }
-function formatTime(iso) {
-  if (!iso) return ''
-  try { return new Date(iso).toLocaleString() } catch { return iso }
-}
-
-// Acting on an example writes the SAME tables the head trains on, so a re-run
-// reflects the correction. Keyed per (concept, list, image); the report ids are
-// frozen at run time, so we just grey out what's been handled in this view.
-const acted = ref({})
-const actedKey = (c, dir, it) => `${c.tag_id}:${dir}:${it.id}`
-const actedLabel = (c, dir, it) => acted.value[actedKey(c, dir, it)] || ''
-
-async function act(c, it, dir, verdict) {
-  const key = actedKey(c, dir, it)
-  let call, label
-  if (dir === 'suggest' && verdict === 'yes') { call = store.applyTag(it.id, c.tag_id); label = 'tagged' }
-  else if (dir === 'suggest' && verdict === 'no') { call = store.rejectTag(it.id, c.tag_id); label = 'rejected' }
-  else if (dir === 'doubts' && verdict === 'no') { call = store.removeTag(it.id, c.tag_id); label = 'removed' }
-  else { call = store.confirmTag(it.id, c.tag_id); label = 'kept' }   // doubt + yes = keep (confirm)
-  try {
-    await call
-    acted.value[key] = label
-  } catch (e) {
-    toast({ text: `Action failed: ${e.message}`, type: 'error' })
-  }
-}
-</script>
-
-<style scoped>
-.fc-muted { color: rgb(var(--v-theme-on-surface-variant)); }
-.fc-cc {
-  padding: 12px 0;
-  border-top: 1px solid rgb(var(--v-theme-surface-light));
-}
-.fc-cc__head { display: flex; align-items: baseline; gap: 8px; margin-bottom: 6px; }
-.fc-cc__name { font-weight: 600; }
-.fc-metrics { width: 100%; max-width: 360px; border-collapse: collapse; font-size: 13px; }
-.fc-metrics th { text-align: right; font-weight: 600; color: rgb(var(--v-theme-on-surface-variant)); padding: 0 8px; }
-.fc-metrics__lbl { text-align: left; }
-.fc-num { text-align: right; font-variant-numeric: tabular-nums; padding: 1px 8px; }
-.fc-win { color: rgb(var(--v-theme-accent)); font-weight: 600; }
-.fc-up { color: rgb(var(--v-theme-success)); }
-.fc-down { color: rgb(var(--v-theme-error)); }
-.fc-curve { margin-bottom: 8px; }
-.fc-curve__pt { margin-left: 10px; font-size: 13px; font-variant-numeric: tabular-nums; }
-.fc-ex__row { margin-top: 8px; }
-.fc-ex__thumbs { display: flex; flex-wrap: wrap; gap: 6px; }
-.fc-ex__item { position: relative; width: 120px; height: 120px; }
-.fc-ex__item--acted { opacity: 0.45; }
-.fc-ex__thumb {
-  display: block; width: 100%; height: 100%; border-radius: 6px;
-  overflow: hidden; background: rgb(var(--v-theme-surface-light));
-  outline: 1px solid transparent; transition: outline-color 0.12s;
-  border: none; padding: 0; cursor: pointer;
-}
-.fc-ex__thumb:hover { outline-color: rgb(var(--v-theme-accent)); }
-.fc-ex__thumb img { width: 100%; height: 100%; object-fit: cover; display: block; }
-.fc-ex__acts { position: absolute; top: 4px; right: 4px; display: flex; gap: 4px; }
-.fc-act {
-  width: 26px; height: 26px; border-radius: 50%; border: none; cursor: pointer;
-  display: flex; align-items: center; justify-content: center; color: #fff;
-  opacity: 0.9; box-shadow: 0 1px 3px rgba(0, 0, 0, 0.4); transition: transform 0.1s;
-}
-.fc-act:hover { opacity: 1; transform: scale(1.1); }
-.fc-act--yes { background: rgb(var(--v-theme-success)); }
-.fc-act--no { background: rgb(var(--v-theme-error)); }
-.fc-ex__badge {
-  position: absolute; bottom: 4px; left: 4px; right: 4px; text-align: center;
-  font-size: 10px; text-transform: uppercase; letter-spacing: 0.05em;
-  background: rgba(0, 0, 0, 0.65); color: #fff; border-radius: 3px; padding: 1px 0;
-}
-</style>

frontend/src/components/settings/TagMaintenanceCard.vue

@@ -42,103 +42,6 @@
     </div>
   </MaintenanceTile>
 
-  <MaintenanceTile
-    icon="mdi-tag-off"
-    title="Legacy migration tags"
-    blurb="Purge retired archive/post/artist + source:* tags."
-    destructive
-  >
-    <p class="fc-muted text-body-2 mb-3">
-      Purge legacy IR-migration tags FC no longer uses: retired/system
-      kinds (<code>archive</code>, <code>post</code>, <code>artist</code> — e.g.
-      <code>BlenderKnight:Hannah_BJ_Loops</code>) plus <code>source:*</code> tags
-      (ImageRepo's old <code>source</code> kind, migrated to <code>general</code>).
-      Provenance and artists are their own systems now, so these are pure noise.
-      Removes them from every image.
-    </p>
-
-    <v-btn
-      color="accent" variant="flat" rounded="pill"
-      prepend-icon="mdi-magnify"
-      :loading="loadingKindPreview"
-      class="mb-3"
-      @click="onKindPreview"
-    >Preview legacy tags</v-btn>
-
-    <div v-if="kindPreview">
-      <p class="text-body-2 mb-2">
-        <strong>{{ kindPreview.count }}</strong> legacy tag(s).
-        <span v-for="(n, k) in kindPreview.by_kind" :key="k" class="fc-muted">
-          {{ k }}: {{ n }}&nbsp;&nbsp;
-        </span>
-        <span v-for="(n, p) in kindPreview.by_prefix" :key="p" class="fc-muted">
-          {{ p }}: {{ n }}&nbsp;&nbsp;
-        </span>
-      </p>
-      <SampleNameGrid
-        v-if="kindPreview.sample_names?.length"
-        :names="kindPreview.sample_names" class="mb-3"
-      />
-      <v-btn
-        color="error" variant="flat" rounded="pill"
-        prepend-icon="mdi-delete-sweep"
-        :disabled="!kindPreview.count"
-        :loading="kindCommitting"
-        @click="onKindCommit"
-      >Delete {{ kindPreview.count }} legacy tag(s)</v-btn>
-    </div>
-  </MaintenanceTile>
-
-  <MaintenanceTile
-    icon="mdi-tag-multiple"
-    title="Reset content tagging"
-    blurb="Delete all general/character tags to re-tag from scratch."
-    destructive
-  >
-    <p class="text-body-2 mb-2">
-      Deletes every <code>general</code> and <code>character</code> tag and
-      removes them from every image, so you can re-tag from scratch with the
-      auto-suggest. <strong>Fandoms and series (with their page order) are
-      kept</strong>, and each image's saved predictions are untouched — open
-      an image and its suggestions reappear.
-    </p>
-    <v-alert type="warning" variant="tonal" density="compact" class="mb-3">
-      Irreversible — there's no undo except restoring a DB backup.
-      Back one up first (Settings → Maintenance → Backup).
-    </v-alert>
-
-    <v-btn
-      color="accent" variant="flat" rounded="pill"
-      prepend-icon="mdi-magnify"
-      :loading="loadingResetPreview"
-      class="mb-3"
-      @click="onResetPreview"
-    >Preview content-tag reset</v-btn>
-
-    <div v-if="resetPreview">
-      <p class="text-body-2 mb-2">
-        <strong>{{ resetPreview.count }}</strong> content tag(s)
-        <span v-for="(n, k) in resetPreview.by_kind" :key="k" class="fc-muted">
-          ({{ k }}: {{ n }})&nbsp;
-        </span>
-        across <strong>{{ resetPreview.applications }}</strong> image
-        application(s).
-      </p>
-      <SampleNameGrid
-        v-if="resetPreview.sample_names?.length"
-        :names="resetPreview.sample_names" class="mb-3"
-      />
-      <v-btn
-        color="error" variant="flat" rounded="pill"
-        prepend-icon="mdi-delete-alert"
-        :disabled="!resetPreview.count"
-        :loading="resetCommitting"
-        @click="onResetCommit"
-      >Delete {{ resetPreview.count }} content tag(s) +
-        {{ resetPreview.applications }} application(s)</v-btn>
-    </div>
-  </MaintenanceTile>
-
   <MaintenanceTile
     icon="mdi-format-letter-case"
     title="Standardize tag casing"
@@ -216,26 +119,6 @@ const {
   emptyPreview: (r) => ({ count: 0, sample_names: r.sample_names || [] }),
 })
 
-// Legacy migration-tag purge.
-const {
-  previewData: kindPreview, previewing: loadingKindPreview,
-  committing: kindCommitting, runPreview: onKindPreview, runCommit: onKindCommit,
-} = usePreviewCommit({
-  preview: () => store.purgeLegacyTags({ dryRun: true }),
-  commit: () => store.purgeLegacyTags({ dryRun: false }),
-  emptyPreview: { count: 0, by_kind: {}, by_prefix: {}, sample_names: [] },
-})
-
-// Reset content tagging (general + character).
-const {
-  previewData: resetPreview, previewing: loadingResetPreview,
-  committing: resetCommitting, runPreview: onResetPreview, runCommit: onResetCommit,
-} = usePreviewCommit({
-  preview: () => store.resetContentTagging({ dryRun: true }),
-  commit: () => store.resetContentTagging({ dryRun: false }),
-  emptyPreview: { count: 0, by_kind: {}, applications: 0, sample_names: [] },
-})
-
 // Standardize casing. The apply DISPATCHES a self-resuming background task (no
 // poll-until-done — that would falsely report complete after the first chunk),
 // so there's no emptyPreview: leave the projection up; a truthy normResult means

frontend/src/components/subscriptions/DownloadsTab.vue

@@ -52,14 +52,9 @@
       {{ String(store.error) }}
     </v-alert>
 
-    <FailingSourcesCard
+    <!-- The failing-sources rollup moved to the Subscriptions landing tab
-      :sources="store.failing"
+         (needs-attention strip, 2026-07-02); the maintenance menu above keeps
-      :retrying-ids="sourcesStore.checkingIds"
+         its bulk-retry via the same shared store action. -->
-      :retrying-all="retryingAll"
-      @retry="onRetrySource"
-      @retry-all="onRetryAll"
-      @view-logs="onViewFailingLogs"
-    />
 
     <div v-if="store.loading && store.events.length === 0" class="fc-dl__loading">
       <v-progress-circular indeterminate color="accent" size="36" />
@@ -124,21 +119,17 @@ import { computed, onMounted, onUnmounted, reactive, ref, watch } from 'vue'
 import { useRoute } from 'vue-router'
 
 import { useDownloadsStore } from '../../stores/downloads.js'
-import { useSourcesStore } from '../../stores/sources.js'
 import DownloadEventRow from '../downloads/DownloadEventRow.vue'
 import DownloadDetailModal from '../downloads/DownloadDetailModal.vue'
 import DownloadStatChips from './DownloadStatChips.vue'
 import DownloadActivitySparkline from './DownloadActivitySparkline.vue'
-import FailingSourcesCard from './FailingSourcesCard.vue'
 import ActiveDownloadsPanel from './ActiveDownloadsPanel.vue'
 import MaintenanceMenu from './MaintenanceMenu.vue'
 import DownloadsFilterPopover from './DownloadsFilterPopover.vue'
 
 const route = useRoute()
 const store = useDownloadsStore()
-const sourcesStore = useSourcesStore()
 const filterModel = ref({ ...store.filter })
-const retryingAll = ref(false)
 
 // Free-text search (client-side over the loaded events) + a toggle to
 // hide "no-change" scheduled scans (status=ok/skipped with 0 files) so
@@ -175,51 +166,18 @@ async function refresh() {
 // rollup + stats so the operator sees it move. Passes force=true so the
 // platform cooldown is bypassed — single-source click is an explicit
 // operator override, useful for rapid auth-fix or fixture testing.
-async function onRetrySource(source) {
+// Bulk retry for the maintenance menu — the shared store action keeps the
-  try {
+// cooldown semantics + tally shape identical to the needs-attention card on
-    await sourcesStore.checkNow(source.id, { force: true })
+// the Subscriptions tab. Toast tallies the three outcomes so the operator
-    toast({ text: `Retry queued for ${source.artist_name || source.platform}`, type: 'success' })
+// can read whether cooldown is the dominant failure mode ("12 deferred
-  } catch (e) {
+// (cooldown)" → yes, rate limit is the issue).
-    if (e?.body?.download_event_id) {
-      toast({ text: 'Already running — see below', type: 'info' })
-    } else {
-      toast({ text: `Retry failed: ${e?.detail || e?.message || e}`, type: 'error' })
-    }
-  } finally {
-    await Promise.all([store.loadFailing(), store.loadStats(24), store.loadFirst()])
-  }
-}
-
-// Bulk retry — leaves cooldown enforcement ON so N failing sources on
-// the same platform don't all retry into the rate limit the cooldown is
-// preventing. Sources deferred by cooldown will be picked up by the
-// next scan tick after the AppSetting expires. Toast tallies the three
-// outcomes so the operator can quickly read whether cooldown is the
-// dominant failure mode ("12 deferred (cooldown)" → yes, rate limit is
-// the issue).
 async function onRetryAll(sources) {
-  retryingAll.value = true
+  const t = await store.retryAllFailing(sources)
-  let ok = 0
+  await store.loadFirst()
-  let conflict = 0
-  let deferred = 0
-  try {
-    for (const s of sources) {
-      try {
-        const body = await sourcesStore.checkNow(s.id)
-        if (body?.status === 'deferred') deferred += 1
-        else ok += 1
-      } catch (e) {
-        if (e?.body?.download_event_id) conflict += 1
-      }
-    }
-  } finally {
-    retryingAll.value = false
-    await Promise.all([store.loadFailing(), store.loadStats(24), store.loadFirst()])
-  }
   const parts = []
-  if (ok) parts.push(`${ok} queued`)
+  if (t.ok) parts.push(`${t.ok} queued`)
-  if (deferred) parts.push(`${deferred} deferred (cooldown)`)
+  if (t.deferred) parts.push(`${t.deferred} deferred (cooldown)`)
-  if (conflict) parts.push(`${conflict} already running`)
+  if (t.conflict) parts.push(`${t.conflict} already running`)
   toast({ text: parts.join(', ') || 'Nothing to retry', type: 'info' })
 }
 
@@ -276,6 +234,16 @@ onMounted(() => {
 })
 onUnmounted(stopPolling)
 
+// The needs-attention card (Subscriptions tab) deep-links here with
+// ?source_id= while this tab may ALREADY be mounted (v-window keeps tabs
+// alive) — react to the query, not just the mount.
+watch(() => route.query.source_id, (v) => {
+  if (v) {
+    filterModel.value = { ...filterModel.value, source_id: Number(v) }
+    refresh()
+  }
+})
+
 // Client-side date filter on the loaded page (avoids a backend round-trip
 // for the date pickers; the existing /api/downloads endpoint can grow
 // these as proper query params later if a UX need shows up).
@@ -366,22 +334,6 @@ async function openDetail(id) {
   await store.loadOne(id)
 }
 
-async function onViewFailingLogs(source) {
-  // Find and open the most recent DownloadEvent for this source.
-  // Reuses the existing DownloadDetailModal — same stdout/stderr/error
-  // surface the row-click in the events feed shows.
-  try {
-    const ev = await store.loadLastForSource(source.id)
-    if (!ev) {
-      toast({
-        text: `No download events recorded for ${source.artist_name || source.platform} yet.`,
-        type: 'warning',
-      })
-    }
-  } catch (e) {
-    toast({ text: `Failed to load logs: ${e.message}`, type: 'error' })
-  }
-}
 </script>
 
 <style scoped>

frontend/src/components/subscriptions/MaintenanceMenu.vue

@@ -18,14 +18,6 @@
         subtitle="Mark stranded pending/running events as error (also runs every 5 min)"
         @click="emit('recover-stalled')"
       />
-      <v-list-item
-        :disabled="true"
-        prepend-icon="mdi-download-box"
-        title="Export failed logs"
-        subtitle="CSV dump — v2"
-      >
-        <v-tooltip activator="parent" location="start">Deferred to a future release</v-tooltip>
-      </v-list-item>
     </v-list>
   </v-menu>
 </template>
@@ -33,7 +25,8 @@
 <script setup>
 // The Downloads tab parent (DownloadsTab.vue) owns the actual retry/sweep
 // handlers — same toast + refresh logic already used by the failing-sources
-// RETRY ALL button. We just emit. Import-pipeline maintenance lives in
+// RETRY ALL button. We just emit. (A permanently-disabled "Export failed
-// Settings → Imports (ImportTaskList.vue), not here.
+// logs" stub sat here until 2026-07-02 — retired; the event list + detail
+// modal cover forensics.)
 const emit = defineEmits(['retry-failed', 'recover-stalled'])
 </script>

frontend/src/components/subscriptions/NeedsAttentionCard.vue

@@ -0,0 +1,73 @@
+<template>
+  <!-- Renders nothing when everything is healthy — the daily answer to
+       "does anything need me?" should be silence, not an empty card. -->
+  <FailingSourcesCard
+    v-if="failing.length"
+    class="mb-4"
+    :sources="failing"
+    :retrying-ids="sourcesStore.checkingIds"
+    :retrying-all="retryingAll"
+    @retry="onRetry"
+    @retry-all="onRetryAll"
+    @view-logs="onViewLogs"
+  />
+</template>
+
+<script setup>
+import { onMounted, onUnmounted, ref } from 'vue'
+import { storeToRefs } from 'pinia'
+import { useRouter } from 'vue-router'
+
+import { toast } from '../../utils/toast.js'
+import FailingSourcesCard from './FailingSourcesCard.vue'
+import { useDownloadsStore } from '../../stores/downloads.js'
+import { useSourcesStore } from '../../stores/sources.js'
+
+// The needs-attention strip on the Subscriptions LANDING tab (2026-07-02):
+// failing sources used to live below the fold of the Downloads tab, so a
+// broken subscription was invisible unless you went looking. Retry logic is
+// shared with the Downloads maintenance menu via the downloads store.
+const store = useDownloadsStore()
+const sourcesStore = useSourcesStore()
+const router = useRouter()
+const { failing } = storeToRefs(store)
+const retryingAll = ref(false)
+let pollId = null
+
+async function onRetry(source) {
+  try {
+    const res = await store.retrySource(source)
+    toast(res === 'queued'
+      ? { text: `Retry queued for ${source.artist_name || source.platform}`, type: 'success' }
+      : { text: 'Already running — check Downloads', type: 'info' })
+  } catch (e) {
+    toast({ text: `Retry failed: ${e?.detail || e?.message || e}`, type: 'error' })
+  }
+}
+
+async function onRetryAll(sources) {
+  retryingAll.value = true
+  try {
+    const t = await store.retryAllFailing(sources)
+    const parts = []
+    if (t.ok) parts.push(`${t.ok} queued`)
+    if (t.deferred) parts.push(`${t.deferred} deferred (cooldown)`)
+    if (t.conflict) parts.push(`${t.conflict} already running`)
+    toast({ text: parts.join(', ') || 'Nothing to retry', type: 'info' })
+  } finally {
+    retryingAll.value = false
+  }
+}
+
+function onViewLogs(source) {
+  // The Downloads tab owns the log/detail surface — deep-link into it
+  // pre-filtered to this source (it watches ?source_id).
+  router.push({ path: '/subscriptions', query: { tab: 'downloads', source_id: source.id } })
+}
+
+onMounted(() => {
+  store.loadFailing()
+  pollId = setInterval(() => { if (!document.hidden) store.loadFailing() }, 60000)
+})
+onUnmounted(() => { if (pollId) clearInterval(pollId) })
+</script>

frontend/src/components/subscriptions/RecentArrivalsCard.vue

@@ -0,0 +1,80 @@
+<template>
+  <v-card v-if="arrivals.length" class="mb-4">
+    <CardHeading icon="mdi-new-box" title="Recent arrivals">
+      <v-spacer />
+      <v-btn
+        variant="text" size="small" rounded="pill"
+        to="/subscriptions?tab=downloads"
+      >
+        All downloads
+        <v-icon end size="small">mdi-arrow-right</v-icon>
+      </v-btn>
+    </CardHeading>
+    <v-card-text class="pt-0">
+      <div
+        v-for="ev in arrivals" :key="ev.id"
+        class="fc-arrival"
+      >
+        <router-link
+          v-if="ev.artist_slug"
+          :to="`/artist/${ev.artist_slug}`"
+          class="fc-arrival__artist"
+        >{{ ev.artist_name || ev.artist_slug }}</router-link>
+        <span v-else class="fc-arrival__artist">{{ ev.artist_name || '—' }}</span>
+        <span class="fc-arrival__meta">
+          {{ ev.platform }} ·
+          {{ ev.files_count ? `${ev.files_count} file(s)` : 'no new files' }}
+          · {{ formatRelative(ev.started_at) }}
+        </span>
+      </div>
+    </v-card-text>
+  </v-card>
+</template>
+
+<script setup>
+import { onMounted, onUnmounted, ref } from 'vue'
+
+import CardHeading from '../common/CardHeading.vue'
+import { formatRelative } from '../../utils/date.js'
+import { useApi } from '../../composables/useApi.js'
+
+// "What came in?" — the other half of the landing tab's daily answer
+// (2026-07-02). Own fetch, own state: the downloads store's event list is
+// the Downloads tab's FILTERED feed; borrowing it would couple this card to
+// whatever filter the operator left that tab on.
+const api = useApi()
+const arrivals = ref([])
+let pollId = null
+
+async function load() {
+  try {
+    const events = await api.get('/api/downloads', {
+      params: { status: 'ok', limit: 25 },
+    })
+    // Real arrivals only — scheduled scans that found nothing are noise here.
+    arrivals.value = events.filter((e) => (e.files_count || 0) > 0).slice(0, 6)
+  } catch { /* non-fatal — the card just hides */ }
+}
+
+onMounted(() => {
+  load()
+  pollId = setInterval(() => { if (!document.hidden) load() }, 60000)
+})
+onUnmounted(() => { if (pollId) clearInterval(pollId) })
+</script>
+
+<style scoped>
+.fc-arrival {
+  display: flex; align-items: baseline; gap: 10px;
+  padding: 4px 0;
+}
+.fc-arrival__artist {
+  font-weight: 600; text-decoration: none;
+  color: rgb(var(--v-theme-accent));
+}
+.fc-arrival__artist:hover { text-decoration: underline; }
+.fc-arrival__meta {
+  font-size: 12px;
+  color: rgb(var(--v-theme-on-surface-variant));
+}
+</style>

frontend/src/components/subscriptions/SettingsTab.vue

@@ -1,12 +1,20 @@
 <template>
   <div>
+    <!-- One extension home (2026-07-02): install/manifest card (moved here
+         from Settings → Overview) + the API key bar it authenticates with.
+         The extension feeds THIS view — cookies + one-click sources — so its
+         setup lives with the rest of the ingestion config. -->
+    <h3 class="fc-section__title">Browser extension</h3>
+    <p class="fc-section__hint">Install, session cookies and the API key it authenticates with.</p>
+    <BrowserExtensionCard class="mb-3" />
     <ExtensionKeyBar class="mb-4" />
 
     <v-alert v-if="credentialsStore.error" type="error" variant="tonal" closable class="mb-4">
       {{ String(credentialsStore.error) }}
     </v-alert>
 
-    <h3 class="text-h6 mb-3">Platform credentials</h3>
+    <h3 class="fc-section__title mt-6">Platform credentials</h3>
+    <p class="fc-section__hint">Per-platform cookies/tokens the downloader uses.</p>
     <v-row>
       <v-col
         v-for="p in platformsStore.list"
@@ -23,7 +31,8 @@
       </v-col>
     </v-row>
 
-    <h3 class="text-h6 mb-3 mt-6">Downloader</h3>
+    <h3 class="fc-section__title mt-6">Downloader</h3>
+    <p class="fc-section__hint">Rate limits and gallery-dl behavior for source checks.</p>
     <v-card variant="outlined">
       <v-card-text v-if="importStore.settings">
         <v-row>
@@ -52,7 +61,8 @@
       </v-card-text>
     </v-card>
 
-    <h3 class="text-h6 mb-3 mt-6">External file-host downloads</h3>
+    <h3 class="fc-section__title mt-6">External file-host downloads</h3>
+    <p class="fc-section__hint">Post-linked mega/gdrive/file-host fetches.</p>
     <v-card variant="outlined">
       <v-card-text v-if="importStore.settings">
         <div class="fc-help mb-2">
@@ -98,7 +108,8 @@
       </v-card-text>
     </v-card>
 
-    <h3 class="text-h6 mb-3 mt-6">Schedule defaults</h3>
+    <h3 class="fc-section__title mt-6">Schedule defaults</h3>
+    <p class="fc-section__hint">Check cadence and failure-badge thresholds for new sources.</p>
     <v-card variant="outlined">
       <v-card-text v-if="importStore.settings">
         <v-row>
@@ -179,6 +190,7 @@ import { onMounted, reactive, ref, watch } from 'vue'
 import { usePlatformsStore } from '../../stores/platforms.js'
 import { useCredentialsStore } from '../../stores/credentials.js'
 import { useImportStore } from '../../stores/import.js'
+import BrowserExtensionCard from '../settings/BrowserExtensionCard.vue'
 import ExtensionKeyBar from '../credentials/ExtensionKeyBar.vue'
 import CredentialUploadDialog from '../credentials/CredentialUploadDialog.vue'
 import CredentialCard from './CredentialCard.vue'
@@ -240,6 +252,21 @@ async function saveDownloader() {
 </script>
 
 <style scoped>
+/* Same section-header language as Settings -> Maintenance/Cleanup (2026-07-02
+   theme unification) so every admin surface reads identically. */
+.fc-section__title {
+  font-size: 0.78rem;
+  font-weight: 700;
+  letter-spacing: 0.06em;
+  text-transform: uppercase;
+  color: rgb(var(--v-theme-accent));
+  margin-bottom: 2px;
+}
+.fc-section__hint {
+  font-size: 0.8rem;
+  color: rgb(var(--v-theme-on-surface-variant));
+  margin-bottom: 12px;
+}
 .fc-help {
   font-size: 12px;
   color: rgb(var(--v-theme-on-surface-variant));

frontend/src/components/subscriptions/SubscriptionsTab.vue

@@ -2,6 +2,12 @@
   <div>
     <SchedulerStatusBar :status="store.scheduleStatus" class="fc-subs__sched" />
 
+    <!-- Daily-use ordering (2026-07-02): what needs me → what came in →
+         the full source list. Both cards render nothing when there's
+         nothing to say. -->
+    <NeedsAttentionCard />
+    <RecentArrivalsCard />
+
     <div class="fc-subs__bar">
       <v-btn color="accent" prepend-icon="mdi-plus" @click="openAddSource(null)">
         Add subscription
@@ -319,6 +325,8 @@ import { useRoute, useRouter } from 'vue-router'
 import { useSourcesStore } from '../../stores/sources.js'
 import { usePlatformsStore } from '../../stores/platforms.js'
 import { useImportStore } from '../../stores/import.js'
+import NeedsAttentionCard from './NeedsAttentionCard.vue'
+import RecentArrivalsCard from './RecentArrivalsCard.vue'
 import SourceRow from './SourceRow.vue'
 import SourceCard from './SourceCard.vue'
 import SourceHealthDot from './SourceHealthDot.vue'

frontend/src/stores/admin.js

@@ -101,12 +101,10 @@ export const useAdminStore = defineStore('admin', () => {
     })
   }
 
-  function purgeLegacyTags(opts = {}) {
+  // Destructive whole-instance reset: deletes ALL general + character tags AND
-    return _dryRunPost('/api/admin/tags/purge-legacy', opts)
+  // their applications (the heads' training data included) — fandom + series
-  }
+  // preserved. dry-run returns a `confirm` token; the apply must pass it back
-
+  // ({ dryRun: false, confirm }) or the server rejects it.
-  // Destructive: deletes ALL general + character tags so the operator can
-  // re-tag from scratch via auto-suggest. fandom + series preserved.
   function resetContentTagging(opts = {}) {
     return _dryRunPost('/api/admin/tags/reset-content', opts)
   }
@@ -154,7 +152,6 @@ export const useAdminStore = defineStore('admin', () => {
     pruneUnusedTags,
     pruneBarePosts,
     reconcileDuplicatePosts,
-    purgeLegacyTags,
     resetContentTagging,
     normalizeTags,
     pollTaskUntilDone,

frontend/src/stores/allowlist.js

@@ -1,44 +0,0 @@
-import { defineStore } from 'pinia'
-import { ref } from 'vue'
-import { useApi } from '../composables/useApi.js'
-
-export const useAllowlistStore = defineStore('allowlist', () => {
-  const api = useApi()
-  const rows = ref([])
-  const loading = ref(false)
-
-  async function load() {
-    loading.value = true
-    try { rows.value = await api.get('/api/allowlist') }
-    finally { loading.value = false }
-  }
-
-  async function updateThreshold(tagId, minConfidence) {
-    await api.patch(`/api/tags/${tagId}/allowlist`, {
-      body: { min_confidence: minConfidence }
-    })
-    const r = rows.value.find(x => x.tag_id === tagId)
-    if (r) {
-      r.min_confidence = minConfidence
-      // The committed threshold changed the covered pool — refresh the row's
-      // coverage so the table stays truthful after a save.
-      try { r.coverage_count = (await coverage(tagId, minConfidence)).count }
-      catch { /* leave the stale count rather than blank it */ }
-    }
-  }
-
-  // Live "at threshold T, a sweep would cover ~N images" projection for the
-  // tuning dashboard. Returns { count, threshold }.
-  async function coverage(tagId, threshold) {
-    return api.get(`/api/tags/${tagId}/allowlist/coverage`, {
-      params: { threshold }
-    })
-  }
-
-  async function remove(tagId) {
-    await api.delete(`/api/tags/${tagId}/allowlist`)
-    rows.value = rows.value.filter(x => x.tag_id !== tagId)
-  }
-
-  return { rows, loading, load, updateThreshold, coverage, remove }
-})

frontend/src/stores/downloads.js

@@ -3,6 +3,7 @@ import { ref } from 'vue'
 import { useApi } from '../composables/useApi.js'
 import { useAsyncAction } from '../composables/useAsyncAction.js'
 import { useInflightToken } from '../composables/useInflightToken.js'
+import { useSourcesStore } from './sources.js'
 
 export const useDownloadsStore = defineStore('downloads', () => {
   const api = useApi()
@@ -106,6 +107,44 @@ export const useDownloadsStore = defineStore('downloads', () => {
     return failing.value
   }
 
+  // --- Failing-source retries (shared by the Subscriptions needs-attention
+  // card and the Downloads maintenance menu — one implementation for both
+  // surfaces). Data-only: callers own the toasts, per this store's style.
+  // A single deliberate retry forces past cooldown; BULK retries keep
+  // cooldown enforcement ON so N failing sources on one platform don't all
+  // retry into the very rate limit the cooldown is preventing.
+  async function retrySource(source) {
+    const sourcesStore = useSourcesStore()
+    try {
+      await sourcesStore.checkNow(source.id, { force: true })
+      return 'queued'
+    } catch (e) {
+      if (e?.body?.download_event_id) return 'already_running'
+      throw e
+    } finally {
+      await Promise.all([loadFailing(), loadStats(24)])
+    }
+  }
+
+  async function retryAllFailing(sources) {
+    const sourcesStore = useSourcesStore()
+    const tally = { ok: 0, deferred: 0, conflict: 0 }
+    try {
+      for (const s of sources) {
+        try {
+          const body = await sourcesStore.checkNow(s.id)
+          if (body?.status === 'deferred') tally.deferred += 1
+          else tally.ok += 1
+        } catch (e) {
+          if (e?.body?.download_event_id) tally.conflict += 1
+        }
+      }
+    } finally {
+      await Promise.all([loadFailing(), loadStats(24)])
+    }
+    return tally
+  }
+
   async function loadActive() {
     const [running, pending] = await Promise.all([
       api.get('/api/downloads', { params: { status: 'running', limit: 50 } }),
@@ -128,5 +167,6 @@ export const useDownloadsStore = defineStore('downloads', () => {
     loadFirst, loadMore, loadOne, loadLastForSource, applyFilter,
     closeDetail, loadStats,
     loadActivity, loadFailing, loadActive, recoverStalled,
+    retrySource, retryAllFailing,
   }
 })

frontend/src/stores/explore.js

@@ -93,13 +93,11 @@ export const useExploreStore = defineStore('explore', () => {
     // a crumb (which snaps the cursor back into the trail — the "loops back"
     // report). Fall back to the full set only if every neighbour's been seen.
     const seen = new Set(breadcrumb.value.map((c) => c.id))
-    let pool = neighbors.value.filter((n) => !seen.has(n.id))
+    const pool = neighbors.value.filter((n) => !seen.has(n.id))
-    if (!pool.length) pool = neighbors.value
+    const cands = pool.length ? pool : neighbors.value
-    // neighbors come similarity-sorted (nearest first). Skip the closest slice —
+    // The list is already pHash-deduped + MMR-diversified server-side (it spans
-    // those near-duplicates are exactly what you get stuck cycling through — and
+    // clusters, not 40 near-dupes), so a plain random pick gives real variance —
-    // pick from the more-varied remainder, for real variance in the walk.
+    // no need to skip the nearest slice the way the raw nearest-list required.
-    const skip = pool.length >= 6 ? Math.floor(pool.length / 3) : 0
-    const cands = pool.slice(skip)
     return cands[Math.floor(Math.random() * cands.length)].id
   }
 

frontend/src/stores/gallery.js

@@ -23,7 +23,7 @@ export const useGalleryStore = defineStore('gallery', () => {
   const error = ref(null)
   const filter = ref({
     tag_ids: [], artist_id: null, media_type: null,
-    sort: 'newest', post_id: null,
+    sort: 'posted_new', post_id: null,
     // #6 structured tag filter (AND-of-OR + exclude). tag_ids are the AND
     // "include" singletons (light editor + back-compat); tag_or is a list of
     // OR-groups (each group ORs, groups AND); tag_exclude is the NOT set.
@@ -154,7 +154,7 @@ export const useGalleryStore = defineStore('gallery', () => {
     if (filter.value.tag_exclude.length) p.tag_not = filter.value.tag_exclude.join(',')
     if (filter.value.artist_id) p.artist_id = filter.value.artist_id
     if (filter.value.media_type) p.media = filter.value.media_type
-    if (filter.value.sort && filter.value.sort !== 'newest') p.sort = filter.value.sort
+    if (filter.value.sort && filter.value.sort !== 'posted_new') p.sort = filter.value.sort
     if (filter.value.platform) p.platform = filter.value.platform
     if (filter.value.untagged) p.untagged = '1'
     if (filter.value.no_artist) p.no_artist = '1'
@@ -191,7 +191,7 @@ export const useGalleryStore = defineStore('gallery', () => {
       tag_exclude: _parseIds(q.tag_not),
       artist_id: _toId(q.artist_id),
       media_type: ['image', 'video'].includes(q.media) ? q.media : null,
-      sort: q.sort === 'oldest' ? 'oldest' : 'newest',
+      sort: ['newest', 'oldest', 'posted_new', 'posted_old'].includes(q.sort) ? q.sort : 'posted_new',
       post_id: _toId(q.post_id),
       platform: q.platform || null,
       untagged: _truthy(q.untagged),
@@ -297,7 +297,7 @@ export function filterToQuery(f) {
   if (f.tag_exclude?.length) q.tag_not = f.tag_exclude.join(',')
   if (f.artist_id) q.artist_id = String(f.artist_id)
   if (f.media_type) q.media = f.media_type
-  if (f.sort && f.sort !== 'newest') q.sort = f.sort
+  if (f.sort && f.sort !== 'posted_new') q.sort = f.sort
   if (f.platform) q.platform = f.platform
   if (f.untagged) q.untagged = '1'
   if (f.no_artist) q.no_artist = '1'