feat(phase 4): vision fetches NC previews; stop writing /data/thumbs

Last consumer of the on-disk thumbnail pipeline was the vision
worker reading /data/thumbs/{id}/medium.webp. Now it asks Nextcloud
for a 640px preview (the same edge size the old thumb used) and
decodes the bytes in-memory — no disk dependency.

- nextcloud_dav.get_preview_bytes: sync sibling of get_preview_async,
  for the celery vision worker (which is sync).
- vision._load_thumb: tries NC preview first; transitional disk
  fallback stays for rows still indexed during the rollout.
- thumbs.WORKER_THUMB_SIZES = set() — generate_thumbnails still runs
  the decode + pHash side-effect (perceptual dedup is mule-only and
  needs original-resolution pixels) but no longer writes thumbnail
  files.

The HTTP thumbnail endpoint's disk fallback path stays in place
unchanged: for NC-404 cases (e.g. iPhone JPEGs mis-extensioned as
.DNG), inline Pillow regeneration still writes a tiny per-photo
file so subsequent requests are fast. That path is rare and the
files are small.

Disk impact: /data/thumbs currently has ~22k medium.webp totaling
~1 GB. They'll stop being read after the worker-vision container
restarts, but no automatic delete — purge with the same find
pattern used for small/large reclaim when ready:

    find /data/thumbs -name "medium.webp" -delete

backend/app/tasks/thumbs.py

@@ -47,13 +47,13 @@ THUMB_SIZES = {
     'large': settings.thumbnails.large
 }
 
-# Sizes the worker actually writes to /data/thumbs. We used to write all
-# three, but the API now proxies Nextcloud's /core/preview for `small`
-# and `large` — only `medium` survives on disk because the vision
-# pipeline (app.tasks.vision) still loads it from there. When vision
-# moves to NC previews too, this set drops to empty and the file
-# pipeline can be deleted entirely.
-WORKER_THUMB_SIZES = {'medium'}
+# Sizes the worker writes to /data/thumbs. Empty set since Phase 4 —
+# the API serves all sizes via Nextcloud's /core/preview proxy, and
+# the vision worker also fetches NC previews on demand instead of
+# reading a local cache. generate_thumbnails still runs the decode-
+# and-pHash side-effect (perceptual dedup is mule-only and needs the
+# original-resolution pixels) but no longer touches the disk.
+WORKER_THUMB_SIZES: set[str] = set()
 
 def get_thumb_path(photo_id: str, size: str, user_id: str = None) -> str:
     """Get the path for a thumbnail file.