Merge pull request 'Modal focus/keyboard polish, Camie-in-autocomplete, re-extract self-resume' (#82) from dev into main

backend/app/services/cleanup_service.py

@@ -12,6 +12,7 @@ the one-and-done GS/IR migration tooling.)
 from __future__ import annotations
 
 import logging
+import time
 from datetime import UTC, datetime, timedelta
 from pathlib import Path
 from typing import Any
@@ -723,7 +724,13 @@ def _reextract_archive_to_post(
         sidecar_path.unlink(missing_ok=True)
 
 
-def reextract_archive_attachments(session: Session, *, images_root: Path) -> dict:
+def reextract_archive_attachments(
+    session: Session,
+    *,
+    images_root: Path,
+    time_budget_seconds: float | None = None,
+    after_id: int = 0,
+) -> dict:
     """Re-process existing PostAttachments that are ACTUALLY archives but were
     filed opaquely before #713 part 1 (extension-only is_archive missed mangled /
     extension-less Patreon attachment names). For each: extract the members,
@@ -731,6 +738,14 @@ def reextract_archive_attachments(session: Session, *, images_root: Path) -> dic
 
     Idempotent — members dedupe by sha256, the archive dedupes by sha — so it's
     safe to run repeatedly. Returns a summary dict for task_run.metadata.
+
+    Time-boxed + resumable: scans PostAttachments in ascending id order starting
+    after ``after_id``. When ``time_budget_seconds`` elapses, stops and reports
+    ``partial=True`` + ``resume_after_id`` (the last scanned id) so the task can
+    re-enqueue itself and continue — a large archive back-catalog can't run the
+    task into the Celery time limit or hog the maintenance lane. A bare re-run
+    (after_id=0) would never advance because an already-extracted archive is
+    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
@@ -742,7 +757,7 @@ def reextract_archive_attachments(session: Session, *, images_root: Path) -> dic
     summary = {
         "scanned": 0, "archives": 0, "members_imported": 0,
         "posts_touched": 0, "skipped_no_post": 0, "skipped_no_artist": 0,
-        "errors": 0,
+        "errors": 0, "partial": False, "resume_after_id": after_id,
     }
     settings = ImportSettings.load_sync(session)
     importer = Importer(
@@ -751,11 +766,15 @@ def reextract_archive_attachments(session: Session, *, images_root: Path) -> dic
     )
 
     attachments = session.execute(
-        select(PostAttachment).order_by(PostAttachment.id)
+        select(PostAttachment)
+        .where(PostAttachment.id > after_id)
+        .order_by(PostAttachment.id)
     ).scalars().all()
     enqueue_ids: list[int] = []
+    start = time.monotonic()
     for att in attachments:
         summary["scanned"] += 1
+        summary["resume_after_id"] = att.id
         stored = Path(att.path)
         try:
             if not stored.is_file() or not is_archive(stored):
@@ -795,6 +814,19 @@ def reextract_archive_attachments(session: Session, *, images_root: Path) -> dic
             summary["posts_touched"] += 1
             enqueue_ids.extend(ids)
 
+        # Time-box the chunk. resume_after_id already points at this attachment,
+        # so the next run starts strictly after it. Checked after the commit so a
+        # half-extracted archive never straddles the boundary.
+        if (
+            time_budget_seconds is not None
+            and time.monotonic() - start >= time_budget_seconds
+        ):
+            summary["partial"] = True
+            break
+    else:
+        # Loop ran to exhaustion — nothing left to resume.
+        summary["partial"] = False
+
     # Thumbnails + ML for the newly-imported members (best-effort; off the
     # critical path — a Redis hiccup must not fail the whole re-extract).
     for img_id in enqueue_ids:

backend/app/tasks/admin.py

@@ -57,6 +57,14 @@ def bulk_delete_images_task(self, *, image_ids: list[int]) -> dict:
         )
 
 
+# Time-box one chunk well under the soft limit so a large archive back-catalog
+# can't run the task into the Celery time limit (or hog the maintenance_long
+# lane). The task re-enqueues itself with the resume cursor until the scan is
+# exhausted — mirrors normalize_tags_task (operator-asked 2026-06-07: reasonable
+# timeout, then re-queue so other work keeps flowing).
+_REEXTRACT_CHUNK_SECONDS = 600
+
+
 @celery.task(
     name="backend.app.tasks.admin.reextract_archive_attachments_task",
     bind=True,
@@ -64,15 +72,30 @@ def bulk_delete_images_task(self, *, image_ids: list[int]) -> dict:
     retry_backoff=15, retry_backoff_max=180, max_retries=1,
     soft_time_limit=1800, time_limit=2400,  # 30 min / 40 min
 )
-def reextract_archive_attachments_task(self) -> dict:
+def reextract_archive_attachments_task(self, after_id: int = 0) -> dict:
     """Wraps cleanup_service.reextract_archive_attachments (#713 part 2):
     re-extract PostAttachments that are actually archives but were filed
-    opaquely before the magic-byte gate, and link their members to the post."""
+    opaquely before the magic-byte gate, and link their members to the post.
+
+    Time-boxed + self-resuming: scans attachments after ``after_id`` and, on a
+    chunk cut, re-enqueues from where it stopped so a big backlog finishes across
+    chunks instead of dying at the soft limit."""
     SessionLocal = _sync_session_factory()
     with SessionLocal() as session:
-        return cleanup_service.reextract_archive_attachments(
+        summary = cleanup_service.reextract_archive_attachments(
             session, images_root=IMAGES_ROOT,
+            time_budget_seconds=_REEXTRACT_CHUNK_SECONDS, after_id=after_id,
         )
+    # More attachments past this chunk's cursor — continue in the next.
+    if summary.get("partial") and summary.get("resume_after_id", 0) > after_id:
+        log.info(
+            "reextract chunk done (%d scanned, %d archives, resume after id %s) "
+            "— re-enqueuing to continue",
+            summary.get("scanned", 0), summary.get("archives", 0),
+            summary["resume_after_id"],
+        )
+        reextract_archive_attachments_task.delay(summary["resume_after_id"])
+    return summary
 
 
 # Time-box one chunk well under the soft limit so a large back-catalog (the

frontend/src/components/modal/FandomPicker.vue

@@ -2,22 +2,37 @@
   <v-card>
     <v-card-title>Pick a fandom</v-card-title>
     <v-card-text>
-      <!-- A2/A3 (2026-06-07): autofocus so the operator types immediately, and
-           picking a fandom (keyboard: type → arrow → Enter, or a click) confirms
-           in one step — selecting IS the decision in this dialog. -->
+      <!-- Keyboard flow (operator-specified 2026-06-07):
+           1. Focus lands here (driven by the dialog's @after-enter calling the
+              exposed focusSearch — `autofocus` is unreliable inside a v-dialog
+              because the focus-trap activates after mount and steals it).
+           2. Tab moves to the "new fandom" field below.
+           3. Selecting a fandom (click, or arrow+Enter to pick from the menu)
+              leaves it in the field; a SECOND Enter (menu closed) accepts it.
+           We bind v-model:menu so the Enter handler can tell "pick from the open
+           menu" (let Vuetify handle) apart from "accept the chosen value". -->
       <v-autocomplete
+        ref="fandomRef"
         v-model="selectedId"
+        v-model:menu="menuOpen"
         :items="store.fandomCache"
         :item-title="(f) => f.name"
         :item-value="(f) => f.id"
         label="Fandom" clearable density="compact"
-        autofocus
-        @update:model-value="onSelect"
+        @keydown.enter="onSearchEnter"
+        @keydown.tab="onSearchTab"
       />
       <v-divider class="my-3" />
       <p class="text-caption mb-2">Or create a new fandom:</p>
       <div class="d-flex" style="gap: 8px;">
-        <v-text-field v-model="newName" placeholder="New fandom name" density="compact" hide-details />
+        <!-- Enter creates; on success focus returns to the dropdown (onCreate),
+             so the operator can immediately Enter-to-accept the new fandom. -->
+        <v-text-field
+          ref="newNameRef"
+          v-model="newName" placeholder="New fandom name"
+          density="compact" hide-details
+          @keydown.enter.prevent="onCreate"
+        />
         <v-btn :disabled="!newName.trim()" rounded="pill" @click="onCreate">Create</v-btn>
       </div>
     </v-card-text>
@@ -34,7 +49,7 @@
 </template>
 
 <script setup>
-import { onMounted, ref } from 'vue'
+import { nextTick, onMounted, ref } from 'vue'
 import { useTagStore } from '../../stores/tags.js'
 
 const emit = defineEmits(['confirm', 'cancel'])
@@ -42,28 +57,61 @@ const store = useTagStore()
 
 const selectedId = ref(null)
 const newName = ref('')
+const menuOpen = ref(false)
+const fandomRef = ref(null)
+const newNameRef = ref(null)
 
 // Always refresh on open so the list reflects fandoms created elsewhere (#712).
 onMounted(() => store.loadFandoms())
 
-async function onCreate() {
-  const f = await store.createFandom(newName.value.trim())
+// Exposed so the parent dialog can focus the search field on @after-enter —
+// the reliable point to grab focus (the dialog transition + focus-trap are done).
+function focusSearch () {
+  nextTick(() => {
+    fandomRef.value?.focus?.()
+    // Keep the menu closed after a programmatic focus so the very next Enter
+    // accepts the value instead of being swallowed as a menu interaction.
+    menuOpen.value = false
+  })
+}
+defineExpose({ focusSearch })
+
+async function onCreate () {
+  const name = newName.value.trim()
+  if (!name) return
+  const f = await store.createFandom(name)
   selectedId.value = f.id
   newName.value = ''
+  // The created fandom is now the selection; hand focus back to the dropdown
+  // so a single Enter accepts it (operator-specified flow 2026-06-07).
+  focusSearch()
 }
-function onConfirm() {
+
+function onConfirm () {
   const f = store.fandomCache.find(x => x.id === selectedId.value)
   if (f) emit('confirm', f)
 }
-// Picking a fandom (keyboard Enter or click) confirms immediately. Ignore the
-// clear action (null) so clearing the field doesn't fire a confirm.
-function onSelect(id) {
-  if (id == null) return
-  onConfirm()
+
+// Enter on the search field. When the menu is open Vuetify is picking the
+// highlighted item (it sets selectedId + closes the menu) — don't also accept
+// on that keystroke. With the menu closed and a value chosen, Enter accepts.
+function onSearchEnter () {
+  if (menuOpen.value) return
+  if (selectedId.value != null) onConfirm()
 }
+
+// Tab from the search field goes to the "new fandom" text field (operator-
+// specified). Shift+Tab keeps normal reverse traversal.
+function onSearchTab (e) {
+  if (e.shiftKey) return
+  e.preventDefault()
+  menuOpen.value = false
+  nextTick(() => newNameRef.value?.focus?.())
+}
+
 // Create the character with no fandom. Emits null so the caller knows this
 // was a deliberate "unassigned", not a cancel.
-function onNoFandom() {
+function onNoFandom () {
   emit('confirm', null)
 }
 </script>

frontend/src/components/modal/TagAutocomplete.vue

@@ -12,44 +12,83 @@
       @keydown.esc="$emit('cancel')"
     />
     <v-list
-      v-if="hits.length || allowCreate"
+      v-if="rows.length"
       ref="listRef"
       density="compact" class="fc-tag-autocomplete__list"
     >
-      <v-list-item
-        v-for="(h, idx) in hits" :key="h.id"
-        :active="idx === highlight" @click="onPick(h)"
-      >
-        <template #prepend>
-          <v-icon size="small" :color="store.colorFor(h.kind)">
-            {{ iconFor(h.kind) }}
-          </v-icon>
-        </template>
-        <v-list-item-title>
-          {{ h.name }}
-          <span v-if="h.fandom_name" class="text-caption">— {{ h.fandom_name }}</span>
-        </v-list-item-title>
-        <template #append>
-          <span class="text-caption">{{ h.kind }}</span>
-        </template>
-      </v-list-item>
-      <v-list-item
-        v-if="allowCreate" :active="highlight === hits.length"
-        @click="onCreate"
-      >
-        <template #prepend>
-          <v-icon size="small" :color="store.colorFor(parsedKind)">
-            {{ iconFor(parsedKind) }}
-          </v-icon>
-        </template>
-        <v-list-item-title>
-          Create "{{ parsedName }}" as {{ parsedKind }}
-        </v-list-item-title>
-      </v-list-item>
+      <template v-for="(row, idx) in rows" :key="row.key">
+        <!-- Existing tag match (server autocomplete). -->
+        <v-list-item
+          v-if="row.type === 'hit'"
+          :active="idx === highlight" @click="onPickRow(row)"
+        >
+          <template #prepend>
+            <v-icon size="small" :color="store.colorFor(row.hit.kind)">
+              {{ iconFor(row.hit.kind) }}
+            </v-icon>
+          </template>
+          <v-list-item-title>
+            {{ row.hit.name }}
+            <span v-if="row.hit.fandom_name" class="text-caption">— {{ row.hit.fandom_name }}</span>
+          </v-list-item-title>
+          <template #append>
+            <span class="text-caption">{{ row.hit.kind }}</span>
+          </template>
+        </v-list-item>
+
+        <!-- ML suggestion for THIS image that matches the typed query. Picking
+             it routes through the same accept path as the Suggestions panel, so
+             it's recorded + drops out of the panel (operator-asked 2026-06-07). -->
+        <v-list-item
+          v-else-if="row.type === 'suggestion'"
+          :active="idx === highlight" @click="onPickRow(row)"
+          class="fc-tag-autocomplete__sugg"
+        >
+          <template #prepend>
+            <v-icon size="small" :color="store.colorFor(row.sugg.category)">
+              {{ iconFor(row.sugg.category) }}
+            </v-icon>
+          </template>
+          <v-list-item-title>
+            {{ row.sugg.display_name }}
+            <span v-if="row.sugg.creates_new_tag" class="text-caption">— new</span>
+          </v-list-item-title>
+          <template #append>
+            <span class="fc-tag-autocomplete__sugg-tag">
+              <v-icon size="x-small">mdi-auto-fix</v-icon>
+              {{ scorePct(row.sugg) }}
+            </span>
+          </template>
+        </v-list-item>
+
+        <!-- Create-new row. -->
+        <v-list-item
+          v-else
+          :active="idx === highlight" @click="onPickRow(row)"
+        >
+          <template #prepend>
+            <v-icon size="small" :color="store.colorFor(parsedKind)">
+              {{ iconFor(parsedKind) }}
+            </v-icon>
+          </template>
+          <v-list-item-title>
+            Create "{{ parsedName }}" as {{ parsedKind }}
+          </v-list-item-title>
+        </v-list-item>
+      </template>
     </v-list>
 
-    <v-dialog v-model="fandomDialog" max-width="480">
-      <FandomPicker @confirm="onFandomChosen" @cancel="fandomDialog = false" />
+    <!-- @after-enter is the reliable moment to focus the picker: the dialog's
+         open transition + focus-trap have settled, so focusSearch lands in the
+         fandom field instead of fighting the trap (operator-flagged 2026-06-07). -->
+    <v-dialog
+      v-model="fandomDialog" max-width="480"
+      @after-enter="fandomPickerRef?.focusSearch?.()"
+    >
+      <FandomPicker
+        ref="fandomPickerRef"
+        @confirm="onFandomChosen" @cancel="onFandomCancel"
+      />
     </v-dialog>
   </div>
 </template>
@@ -57,10 +96,15 @@
 <script setup>
 import { computed, nextTick, onMounted, ref, watch } from 'vue'
 import { useTagStore } from '../../stores/tags.js'
+import { useSuggestionsStore } from '../../stores/suggestions.js'
 import FandomPicker from './FandomPicker.vue'
 
-const emit = defineEmits(['pick-existing', 'pick-new', 'cancel'])
+const emit = defineEmits(['pick-existing', 'pick-new', 'accept-suggestion', 'cancel'])
 const store = useTagStore()
+// The image's ML (Camie) suggestions, surfaced inline in this dropdown so the
+// operator can pick a suggestion while typing instead of hunting for it in the
+// Suggestions panel below (operator-asked 2026-06-07).
+const suggestions = useSuggestionsStore()
 
 // Autofocus on modal open so the operator can type the moment the view
 // modal renders, no extra click required (operator-asked 2026-06-01).
@@ -74,10 +118,11 @@ const store = useTagStore()
 // the modal's stacked-layout breakpoint (ImageViewer.vue). matchMedia is
 // safe on plain HTTP (not a secure-context API).
 const inputRef = ref(null)
-onMounted(() => {
+function focusInput () {
   if (window.matchMedia?.('(max-width: 900px)')?.matches) return
   nextTick(() => inputRef.value?.focus?.())
-})
+}
+onMounted(focusInput)
 
 // Single text input; no kind dropdown. Client-side mirror of the
 // backend's parse_kind_prefix lives below — kept in sync with
@@ -89,6 +134,7 @@ const hits = ref([])
 const highlight = ref(0)
 const listRef = ref(null)
 const fandomDialog = ref(false)
+const fandomPickerRef = ref(null)
 let pendingNewName = null
 
 const KNOWN_KINDS = new Set([
@@ -136,8 +182,42 @@ const allowCreate = computed(() => {
   )
 })
 
+function scorePct (s) { return `${Math.round(s.score * 100)}%` }
+
+// This image's suggestions that match the typed query, minus any the server
+// autocomplete already returned (same name+kind) so a tag never shows twice.
+const suggestionHits = computed(() => {
+  const q = parsedName.value.toLowerCase()
+  if (!q) return []
+  const seen = new Set(hits.value.map(h => `${h.kind}:${h.name.toLowerCase()}`))
+  const out = []
+  for (const list of Object.values(suggestions.byCategory)) {
+    for (const s of list || []) {
+      const key = `${s.category}:${s.display_name.toLowerCase()}`
+      if (!s.display_name.toLowerCase().includes(q)) continue
+      if (seen.has(key)) continue
+      seen.add(key)
+      out.push(s)
+    }
+  }
+  // Best matches first; cap so the dropdown stays scannable.
+  out.sort((a, b) => b.score - a.score)
+  return out.slice(0, 6)
+})
+
+// One ordered list backing both the rendered dropdown and keyboard nav, so the
+// highlight index maps 1:1 to a row regardless of which section it's in.
+const rows = computed(() => {
+  const r = hits.value.map(h => ({ type: 'hit', key: `h${h.id}`, hit: h }))
+  for (const s of suggestionHits.value) {
+    r.push({ type: 'suggestion', key: `s${s.category}:${s.display_name}`, sugg: s })
+  }
+  if (allowCreate.value) r.push({ type: 'create', key: 'create' })
+  return r
+})
+
 function moveHighlight (delta) {
-  const total = hits.value.length + (allowCreate.value ? 1 : 0)
+  const total = rows.value.length
   if (total === 0) return
   highlight.value = (highlight.value + delta + total) % total
   // Keep the highlighted row visible — the list is capped at 240px and arrowing
@@ -150,7 +230,18 @@ function moveHighlight (delta) {
   })
 }
 
-function onPick (hit) { emit('pick-existing', hit); reset() }
+// Dispatch a chosen dropdown row by its type.
+function onPickRow (row) {
+  if (row.type === 'hit') { emit('pick-existing', row.hit); reset() }
+  else if (row.type === 'suggestion') { onPickSuggestion(row.sugg) }
+  else { onCreate() }
+}
+
+// Picking an ML suggestion hands it to the parent, which runs the same accept
+// flow as the Suggestions panel (creates the tag if raw, records acceptance,
+// drops it from the panel). reset() clears the query; the panel-drop makes it
+// fall out of suggestionHits reactively.
+function onPickSuggestion (s) { emit('accept-suggestion', s); reset() }
 
 function onCreate () {
   const name = parsedName.value
@@ -178,21 +269,29 @@ function onFandomChosen (fandom) {
   })
   pendingNewName = null
   reset()
+  // Return focus to the tag input so the keyboard flow continues straight into
+  // the next tag instead of dropping to <body> (operator-flagged 2026-06-07).
+  focusInput()
+}
+
+// Cancelling the fandom dialog drops the pending character and hands focus back
+// to the tag input, same as a successful pick.
+function onFandomCancel () {
+  fandomDialog.value = false
+  pendingNewName = null
+  focusInput()
 }
 
 function onEnter () {
-  if (highlight.value < hits.value.length) {
-    onPick(hits.value[highlight.value])
-  } else if (allowCreate.value) {
-    onCreate()
-  }
+  const row = rows.value[highlight.value]
+  if (row) onPickRow(row)
 }
 
-// B5 (2026-06-07): when the suggestion list is open, Tab accepts the
-// highlighted row (standard autocomplete convention) instead of leaving the
-// field. With the list closed it falls through to normal focus traversal.
+// B5 (2026-06-07): when the dropdown is open, Tab accepts the highlighted row
+// (standard autocomplete convention) instead of leaving the field. With the
+// list closed it falls through to normal focus traversal.
 function onTab (e) {
-  if (hits.value.length || allowCreate.value) {
+  if (rows.value.length) {
     e.preventDefault()
     onEnter()
   }
@@ -210,4 +309,14 @@ function reset () { query.value = ''; hits.value = []; highlight.value = 0 }
   border: 1px solid rgb(var(--v-theme-surface-light));
   border-radius: 6px;
 }
+/* Mark ML-suggestion rows so they read as distinct from typed/known matches. */
+.fc-tag-autocomplete__sugg {
+  border-left: 2px solid rgb(var(--v-theme-accent), 0.5);
+}
+.fc-tag-autocomplete__sugg-tag {
+  display: inline-flex; align-items: center; gap: 2px;
+  font-size: 11px;
+  font-family: 'JetBrains Mono', monospace;
+  color: rgb(var(--v-theme-accent));
+}
 </style>

frontend/src/components/modal/TagPanel.vue

@@ -14,6 +14,7 @@
 
     <TagAutocomplete
       @pick-existing="onPickExisting" @pick-new="onPickNew"
+      @accept-suggestion="onAcceptSuggestion"
     />
 
     <v-alert v-if="errorMsg" type="error" variant="tonal" class="mt-2" closable>
@@ -45,6 +46,7 @@
 <script setup>
 import { ref } from 'vue'
 import { useModalStore } from '../../stores/modal.js'
+import { useSuggestionsStore } from '../../stores/suggestions.js'
 import TagChip from './TagChip.vue'
 import TagAutocomplete from './TagAutocomplete.vue'
 import SuggestionsPanel from './SuggestionsPanel.vue'
@@ -52,6 +54,7 @@ import TagRenameDialog from './TagRenameDialog.vue'
 import FandomSetDialog from './FandomSetDialog.vue'
 
 const modal = useModalStore()
+const suggestions = useSuggestionsStore()
 const errorMsg = ref(null)
 
 async function onRemove(tagId) {
@@ -69,6 +72,16 @@ async function onPickNew(payload) {
   try { await modal.createAndAdd(payload) }
   catch (e) { errorMsg.value = e.message }
 }
+// A suggestion picked from the autocomplete dropdown runs the SAME path as the
+// Suggestions panel's Accept: the store creates the tag if it's raw, records the
+// acceptance, and drops it from the panel; then we refresh the chip rail.
+async function onAcceptSuggestion(s) {
+  errorMsg.value = null
+  try {
+    await suggestions.accept(s)
+    await modal.reloadTags()
+  } catch (e) { errorMsg.value = e.message }
+}
 
 const renameDialog = ref(false)
 const renameTarget = ref(null)

tests/test_reextract_archives.py

@@ -6,7 +6,7 @@ import io
 import zipfile
 
 import pytest
-from PIL import Image
+from PIL import Image, ImageDraw
 from sqlalchemy import select
 
 from backend.app.models import (
@@ -28,6 +28,18 @@ def _jpeg(color, size=256):
     return buf.getvalue()
 
 
+def _patterned_jpeg(idx, size=256):
+    """Structurally distinct per idx so the members don't phash-collapse — solid
+    colors collide to distance 0 and the second would dedupe away (the two-axis
+    dedup gotcha). A rectangle at an idx-derived position keeps phashes apart."""
+    img = Image.new("RGB", (size, size), "white")
+    x = 20 + idx * 60
+    ImageDraw.Draw(img).rectangle([x, x, x + 70, x + 70], fill="black")
+    buf = io.BytesIO()
+    img.save(buf, "JPEG")
+    return buf.getvalue()
+
+
 def test_reextract_links_archive_members_to_post(db_sync, tmp_path, monkeypatch):
     from backend.app.tasks import ml as ml_mod
     from backend.app.tasks import thumbnail as thumb_mod
@@ -89,3 +101,67 @@ def test_reextract_links_archive_members_to_post(db_sync, tmp_path, monkeypatch)
     )
     assert again["members_imported"] == 0
     assert db_sync.execute(select(ImageRecord)).scalars().all() == images
+
+
+def test_reextract_timebox_resumes_from_cursor(db_sync, tmp_path, monkeypatch):
+    """A 0-second budget cuts the chunk after the first attachment and reports a
+    resume cursor; the next run starts strictly after it and finishes the rest."""
+    from backend.app.tasks import ml as ml_mod
+    from backend.app.tasks import thumbnail as thumb_mod
+
+    monkeypatch.setattr(thumb_mod.generate_thumbnail, "delay", lambda *a, **k: None)
+    monkeypatch.setattr(ml_mod.tag_and_embed, "delay", lambda *a, **k: None)
+
+    images_root = tmp_path / "images"
+    images_root.mkdir()
+
+    artist = Artist(name="Bob", slug="bob")
+    db_sync.add(artist)
+    db_sync.flush()
+    source = Source(
+        artist_id=artist.id, platform="patreon",
+        url="https://patreon.com/bob", enabled=True, config_overrides={},
+    )
+    db_sync.add(source)
+    db_sync.flush()
+    post = Post(
+        source_id=source.id, artist_id=artist.id, external_post_id="42",
+        post_url="https://www.patreon.com/posts/42",
+    )
+    db_sync.add(post)
+    db_sync.flush()
+
+    store_dir = images_root / "attachments" / "two"
+    store_dir.mkdir(parents=True)
+    att_ids = []
+    for n in range(2):
+        arc = store_dir / f"{n}_archive"
+        with zipfile.ZipFile(arc, "w") as zf:
+            zf.writestr(f"member{n}.jpg", _patterned_jpeg(n))
+        sha = hashlib.sha256(arc.read_bytes()).hexdigest()
+        att = PostAttachment(
+            post_id=post.id, artist_id=artist.id, sha256=sha, path=str(arc),
+            original_filename=arc.name, ext="", size_bytes=arc.stat().st_size,
+        )
+        db_sync.add(att)
+        db_sync.flush()
+        att_ids.append(att.id)
+    db_sync.commit()
+
+    # Budget 0 → break right after the first attachment commits.
+    first = cleanup_service.reextract_archive_attachments(
+        db_sync, images_root=images_root, time_budget_seconds=0.0,
+    )
+    assert first["partial"] is True
+    assert first["scanned"] == 1
+    assert first["members_imported"] == 1
+    assert first["resume_after_id"] == att_ids[0]
+
+    # Resume strictly after the cursor — picks up the second, then runs dry.
+    second = cleanup_service.reextract_archive_attachments(
+        db_sync, images_root=images_root, after_id=att_ids[0],
+    )
+    assert second["partial"] is False
+    assert second["scanned"] == 1
+    assert second["members_imported"] == 1
+    assert len(db_sync.execute(select(ImageRecord)).scalars().all()) == 2