Library sync on folder change

[gosku]

Sync the Library when a folder is added or updated

What

The Library page now imports images the moment you add a folder or change a folder's path, instead of only at make start. Each folder shows its sync progress live, and it works in both install modes (lite: SQLite, no broker; full: PostgreSQL + Celery).

Removing a folder is unchanged: it only stops monitoring and never deletes images (add-only, per ADR 010).

Why

Until now, adding or repointing a folder on the Library page only updated the monitored list. Nothing was imported until the next restart, so a user could register a folder full of photos and see nothing happen, with no feedback and no obvious reason. This closes that gap: you register or repoint a folder and it syncs right away, with visible progress, while the page stays responsive even for a first import of tens of thousands of files.

How

• Trigger. The add and path-update views call a trigger_folder_sync use case after the folder change; removal does nothing.
• Execution strategy lives in the application layer, not the view.
  • Full mode: enqueue one Celery task per new image and return immediately.
  • Lite mode: run the sync in a background daemon thread (via a small services/background.py runner) so the request never blocks. The thread lives in the web process and outlives the request, so navigating away does not stop it.
• Progress. A new SyncRun model records state (scanning, processing, completed, failed, interrupted), counters, and timestamps, persisted so any request can read it. The folder row polls a status endpoint over HTMX and shows Scanning..., a progress bar, then a summary like Imported 36, skipped 3. A conditional unique constraint allows at most one active run per folder (which doubles as the "already syncing" guard), and counters use atomic increments with a race-safe finalize so concurrent Celery workers stay consistent.
• Concern separation. The pure process_image operation and the generic image-processing task (shared with the manual import command) are untouched. A process_synced_image use case composes process_image with the progress bookkeeping; both the new Celery task and the lite thread delegate to it.
• Single-folder scope with recovery. The per-folder sync logic is factored out; the startup sync_library loops over it and first marks any dangling run interrupted, so a run killed mid-flight self-heals on the next start.
• SQLite tuning (lite). WAL and a busy timeout are enabled from the engine, so the background writer never blocks foreground reads and a colliding write waits rather than failing with "database is locked".

Notes

• Full mode: restart the Celery worker after deploying so it registers the new task. Until then those messages are discarded and the next make start re-syncs.
• Documentation updated (docs/library_sync.md) and a new decision record added (ADR 011).
• Implemented layer by layer (data, domain, application, interface) with tests at every layer.