- _assess_via_orch(): uses CFOrchClient.task_allocate('snipe', 'image_assessment')
with multimodal image_url payload; falls back to local LLMRouter on TaskNotFound
- _assess_via_local_llm(): lazy LLMRouter import, unchanged local path
- CF_ORCH_URL env var selects path at runtime; local users unaffected
- Added circuitforge-orch>=0.1.0 to main dependencies
- 5 new runner tests covering orch happy path, task tag, image_url payload format,
TaskNotFound fallback, and non-JSON response handling (13 tests total, 244 suite)
- Rename 002_background_tasks.sql → 007_background_tasks.sql to avoid
collision with existing 002_add_listing_format.sql migration
- Add CREATE UNIQUE INDEX on trust_scores(listing_id) in same migration
so save_trust_scores() can use ON CONFLICT upsert semantics
- Add Store.save_trust_scores() — upserts scores keyed by listing_id;
preserves photo_analysis_json so runner writes are never clobbered
- runner.py: replace raw sqlite3.connect() with get_connection() throughout
(timeout=30 + WAL mode); fix connection leak in insert_task via try/finally
- _run_trust_photo_analysis: read 'user_db' from params to write results to
the correct per-user DB in cloud mode (was silently writing to wrong DB)
- main.py lifespan: use _shared_db_path() in cloud mode so background_tasks
queue lives in shared DB, not _LOCAL_SNIPE_DB
- Add _enqueue_vision_tasks() and call it after score_batch() — this is the
missing enqueue call site; gated by features.photo_analysis (Paid tier)
- Test fixture: add missing 'stage' column to background_tasks schema
