feat(system-tags): process vs chrome groups + WIP provisional auto-apply (#1464)

Backend for the system-tag behavior refactor (milestone #157). editor screenshot
moves from chrome (hidden) to the PROCESS group (shown, like wip); wip+editor gain
provisional auto-apply so they stop needing endless manual identification —
without a runaway loop.

- tag.py: split PRESENTATION_SYSTEM_TAGS → CHROME_SYSTEM_TAGS (banner) +
  PROCESS_SYSTEM_TAGS (wip, editor screenshot).
- heads.py: generalize presentation_auto_apply_sweep → system_tag_auto_apply_sweep
  (mode chrome|process). Same Guard 1 (skip human/confirmed) + Guard 2 (ring-loud
  conflict → PresentationReview). process mode uses source 'process_auto' and does
  NOT hide (hide is a gallery-query effect of group membership).
- training_data._AUTO_SOURCES += 'process_auto' → the head never trains on its own
  auto-applied output; only wip_title/manual train it (the runaway break).
- ml_settings: process_auto_apply_enabled (OFF, opt-in) + threshold + conflict
  threshold. presentation_review.mode ('chrome'|'process'). Migration 0086.
- gallery_service: default-hide reads CHROME only (editor now shows); Explore
  neighbors exclude the whole PROCESS group.
- tasks/ml + celery beat: scheduled_process_auto_apply (daily, opt-in); prune
  covers both modes.
- api: ml_admin process_* CRUD+validation; hidden-review returns mode.
- tests: rename chrome sweep calls; new test_process_auto_apply (apply, guards,
  mode flag, no-self-train); gallery test asserts editor now visible.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

backend/app/services/ml/heads.py

@@ -39,7 +39,7 @@ from ...models import (
     TagPositiveConfirmation,
     TagSuggestionRejection,
 )
-from ...models.tag import PRESENTATION_SYSTEM_TAGS, image_tag
+from ...models.tag import CHROME_SYSTEM_TAGS, PROCESS_SYSTEM_TAGS, image_tag
 from .training_data import (
     _AUTO_SOURCES,
     _auto_apply_point,
@@ -759,18 +759,42 @@ def auto_apply_sweep(
 
 
 _PRESENTATION_SOURCE = "presentation_auto"
+_PROCESS_SOURCE = "process_auto"
+
+# System-tag auto-apply modes (#1464). Both modes run the identical sweep — apply
+# a system tag at a flat threshold with a PROVISIONAL source + a ring-loud review
+# guard — and differ ONLY in which tags, which settings knobs, and which
+# source/review-mode. 'chrome' (banner) is HIDDEN from the gallery; 'process'
+# (wip / editor screenshot) stays VISIBLE (the hide is a gallery-query effect of
+# the tag's group membership, not of this sweep).
+_SWEEP_MODES = {
+    "chrome": {
+        "names": CHROME_SYSTEM_TAGS,
+        "enabled": "presentation_auto_apply_enabled",
+        "threshold": "presentation_auto_apply_threshold",
+        "conflict": "presentation_conflict_threshold",
+        "source": _PRESENTATION_SOURCE,
+    },
+    "process": {
+        "names": PROCESS_SYSTEM_TAGS,
+        "enabled": "process_auto_apply_enabled",
+        "threshold": "process_auto_apply_threshold",
+        "conflict": "process_conflict_threshold",
+        "source": _PROCESS_SOURCE,
+    },
+}
 
 
-def _presentation_heads(session: Session, embedding_version: str):
-    """Trained heads for the presentation chrome tags (banner / editor screenshot).
-    They fire at the FLAT presentation threshold regardless of graduation — a head
-    exists once the operator has labelled enough chrome (head_min_positives)."""
+def _system_tag_heads(session: Session, embedding_version: str, names):
+    """Trained heads for a system-tag group (chrome banner / process wip+editor).
+    They fire at the group's FLAT threshold regardless of graduation — a head
+    exists once the operator has labelled enough (head_min_positives)."""
     return session.execute(
         select(TagHead.tag_id, Tag.name, TagHead.weights, TagHead.bias)
         .join(Tag, Tag.id == TagHead.tag_id)
         .where(TagHead.embedding_version == embedding_version)
         .where(Tag.is_system.is_(True))
-        .where(Tag.name.in_(PRESENTATION_SYSTEM_TAGS))
+        .where(Tag.name.in_(names))
     ).all()
 
 
@@ -802,27 +826,33 @@ def _valued_image_ids(session: Session) -> set[int]:
     return {r[0] for r in rows}
 
 
-def presentation_auto_apply_sweep(session: Session, dry_run: bool = False) -> dict:
-    """Auto-hide presentation chrome (banner / editor screenshot) at the FLAT
-    presentation threshold (#141) — NOT the per-head graduated threshold. Two
-    guards keep it safe: (1) never hide an image carrying a human/confirmed content
-    tag; (2) if an image about to be hidden ALSO scores >= the conflict threshold
-    on a content head, still hide it but flag it (PresentationReview) so the Hidden
-    view surfaces "also looks like <X>" for review. No-op unless
-    presentation_auto_apply_enabled. numpy-only (no sklearn). Returns
-    {n_applied, n_flagged, concepts}."""
+def system_tag_auto_apply_sweep(
+    session: Session, *, mode: str, dry_run: bool = False
+) -> dict:
+    """Auto-apply a system-tag group at its FLAT threshold. mode='chrome' (banner,
+    #141) hides the image; mode='process' (wip / editor screenshot, #1464) keeps it
+    VISIBLE — the ONLY difference is the tag group's gallery membership, not this
+    sweep. Two guards keep it safe: (1) never touch an image carrying a
+    human/confirmed content tag; (2) if the image ALSO scores >= the conflict
+    threshold on a content head, still apply but flag it (PresentationReview,
+    mode=<mode>) so the review strip surfaces "also looks like <X>". The source is
+    PROVISIONAL so the head never trains on its own output. No-op unless the mode's
+    enabled flag is set. numpy-only (no sklearn). Returns {n_applied, n_flagged,
+    concepts}."""
     import numpy as np
     from sqlalchemy.dialects.postgresql import insert as pg_insert
 
+    cfg = _SWEEP_MODES[mode]
     settings = _settings(session)
-    if not dry_run and not settings.presentation_auto_apply_enabled:
+    if not dry_run and not getattr(settings, cfg["enabled"]):
         return {"n_applied": 0, "n_flagged": 0, "concepts": []}
     ver = settings.embedder_model_version
-    pres = _presentation_heads(session, ver)
+    pres = _system_tag_heads(session, ver, cfg["names"])
     if not pres:
         return {"n_applied": 0, "n_flagged": 0, "concepts": []}
-    thr = float(settings.presentation_auto_apply_threshold)
-    conflict_thr = float(settings.presentation_conflict_threshold)
+    thr = float(getattr(settings, cfg["threshold"]))
+    conflict_thr = float(getattr(settings, cfg["conflict"]))
+    source = cfg["source"]
 
     Wp = np.vstack([np.asarray(r.weights, dtype=np.float32) for r in pres])
     bp = np.asarray([r.bias for r in pres], dtype=np.float32)
@@ -884,11 +914,13 @@ def presentation_auto_apply_sweep(session: Session, dry_run: bool = False) -> di
                         pg_insert(image_tag)
                         .values(
                             image_record_id=iid, tag_id=tid,
-                            source=_PRESENTATION_SOURCE,
+                            source=source,
                         )
                         .on_conflict_do_nothing()
                     )
-                # Guard 2: also looks like content → hide but flag for review.
+                # Guard 2: also looks like real content → still apply, but flag it
+                # for the review strip instead of silently marking (chrome hides,
+                # process stays visible — either way the operator gets a heads-up).
                 if Wc is not None and float(max_c[idx]) >= conflict_thr:
                     n_flagged += 1
                     if not dry_run:
@@ -898,6 +930,7 @@ def presentation_auto_apply_sweep(session: Session, dry_run: bool = False) -> di
                                 image_record_id=iid, tag_id=tid,
                                 conflict_tag_id=conf_tag_ids[int(arg_c[idx])],
                                 conflict_score=float(max_c[idx]),
+                                mode=mode,
                             )
                             .on_conflict_do_nothing()
                         )

backend/app/services/ml/training_data.py

@@ -29,7 +29,10 @@ from ...models.tag import image_tag
 # a CCIP reference) unless the operator confirms them (milestone 139). Keeping
 # auto-applied predictions out of training is what makes them "soft" — a misfire
 # can't reinforce itself, so the retraction sweep can actually drop it.
-_AUTO_SOURCES = ("head_auto", "ccip_auto", "ml_auto", "presentation_auto")
+# `process_auto` (#1464): wip/editor screenshot applied by the process sweep are
+# ALSO provisional — the head must learn only from title (`wip_title`) + manual
+# labels, never its own auto-applied output, or it would runaway (operator 2026-07-12).
+_AUTO_SOURCES = ("head_auto", "ccip_auto", "ml_auto", "presentation_auto", "process_auto")
 
 
 def _hygiene_excluded_ids(session: Session) -> set[int]: