diff --git a/containers/120-mule-images.md b/containers/120-mule-images.md
index ad12e2e..3a005d6 100644
--- a/containers/120-mule-images.md
+++ b/containers/120-mule-images.md
@@ -18,11 +18,12 @@ Hosts `mule-image` / "mulita" — the photos app at `photos.hubris.network`. Aut
 | ----------------- | ------ | ----- |
 | frontend          | 3000   | Reverse-proxied by Caddy |
 | backend           | 8001   | FastAPI |
-| worker-vision     | —      | ML scan worker |
-| worker-light      | —      | Lightweight worker (+ celery beat scheduler since 2026-05-11) |
-| db                | (pg)   | pgvector |
+| worker-light      | —      | Sole celery worker — runs default/high/low queues + beat scheduler |
+| db                | (pg)   | postgres:16 (was pgvector/pg16 until 2026-05-14) |
 | redis             | (rd)   | queue |
 
+**No vision / ML services.** The OpenCLIP ONNX classifier and its dedicated `worker-vision` were ripped out on 2026-05-14 — see Changelog.
+
 `worker-watcher` was retired 2026-05-11 — file events come from NC webhook_listeners instead of watchfiles inotify. See "Nextcloud integration" below.
 
 `.env` is **untracked** — `git checkout .env` will wipe it. Holds:
@@ -90,6 +91,50 @@ For pushes from inside the LXC, gitea creds at `/etc/mule-deploy/git-credentials
 
 ## Changelog
 
+### 2026-05-14 — Drop AI/vision pipeline, plain Postgres, DB↔FS refresh
+
+**AI removal (commits `a27267f` + `423a73a`).** The OpenCLIP-on-ONNX
+binary classifier (photography vs other) and all its scaffolding gone:
+
+- Backend: `app/services/vision/`, `app/tasks/vision.py`,
+  `app/services/feature_flags.py`, `app/routers/features.py` deleted.
+  Admin `/admin/feature-flags`, `/admin/ai/{backfill,rescan}` endpoints
+  removed. `Photo.needs_review` column + `ix_photos_needs_review`
+  index dropped (alembic `0019_drop_ai_remnants`).
+- Frontend: AI Settings tab, `useFeaturesQuery`, "Needs Review" sidebar
+  entry + filter, `needs_review` URL param plumbing, FeatureFlag types
+  all gone.
+- Infra: `worker-vision` compose service + `models_data` volume
+  deleted. `worker-light` no longer runs `python -m
+  app.services.vision.bootstrap_models` before celery. Backend
+  Dockerfile drops the dedicated torch RUN layer. `requirements.txt`
+  drops `torch`, `torchvision`, `open-clip-torch`, `onnxruntime`.
+
+**Postgres image swap.** `pgvector/pgvector:pg16` → `postgres:16`. The
+0019 migration drops the (now-unused) `vector` extension first; the
+SQL was pre-applied via `psql` against the still-pgvector container
+and `alembic_version` stamped to `0019` so the new backend's
+`bootstrap.py` upgrade-head was a no-op. After the swap surfaced a
+collation-version mismatch (Bookworm glibc 2.36 → Trixie glibc 2.41),
+the fix was `REINDEX DATABASE mulita` + `ALTER DATABASE … REFRESH
+COLLATION VERSION` on `mulita`, `postgres`, `template1`. The data
+volume was preserved across the image change.
+
+**One-shot DB ↔ filesystem refresh.** New script
+`backend/scripts/full_refresh.py` (run as
+`docker exec -w /app mulita-backend python -m scripts.full_refresh
+[--dry-run]`). Phases: `cleanup_data_integrity` → inline scan of every
+active SourceRoot → `prune_missing_photos(dry_run=False)` → new
+`prune_orphan_thumbnails` helper that removes
+`/data/thumbs/{user}/{photo}/` dirs for photo_ids that left the DB.
+First run: 0 missing photos, 1 stale folder row, **7982 orphan
+thumbnail dirs** reaped.
+
+**Manual cleanup post-deploy.** `docker compose up -d --build
+--force-recreate` doesn't reap services removed from the YAML, so
+`mulita-worker-vision` had to be `docker rm -f`'d by hand. The
+`mule-image_models_data` volume was likewise removed manually.
+
 ### 2026-05-11 — Stop duplicating Nextcloud's indexing (Phase 1 + 2)
 
 Two big architectural shifts in one session, both aimed at killing work mule was doing that NC already does for the same source files.