chore(dedupe): collapse low-risk redundancies from code review

[lstein]

Summary

Low-risk batch of redundancy fixes from the code review. Net +172 / −263 across 6 files.

Backend:

• Removed the duplicate /filetree/home endpoint from routers/album.py — routers/filetree.py already serves the canonical one. Whichever router was included last had been silently winning.
• Extracted _compute_curation() and _validate_curation_request(). /curate (async via background task) and /curate_sync were carrying ~150 lines of identical Monte Carlo / mapping / selection logic. The async path passes a tiny on_iteration callback to drive its ProgressTracker.
• Routed delete_image, move_images, copy_images in routers/index.py through the existing get_embeddings_for_album helper instead of three near-identical inline Embeddings(embeddings_path=..., encoder_spec=...) constructions.
• Added _l2_normalize(x, axis, eps=1e-12) in embeddings.py and used it at all three numpy-normalize sites — also fixes the no-epsilon NaN risk in find_duplicate_clusters that the review flagged.
• Added _normalized_filtered_embeddings() shared loader for FPS / KMeans. get_fps_indices_global and get_kmeans_indices_global shared a ~30-line "open npz / build valid_mask / normalize" preamble — that part is now in one place.

Frontend:

• Deleted the duplicate createPathField in album-manager.js (defined twice, byte-for-byte identical — the second shadowed the first).
• Replaced new ScoreDisplay() in seek-slider.js with the exported scoreDisplay singleton from score-display.js. The second instance was double-binding the star click handler on #scoreText, so bookmark toggles fired twice and immediately cancelled themselves.

Deferred to follow-up PRs

The bigger structural items from the review (FastAPI Depends() for validate_album_exists/get_embeddings_for_album/check_album_lock, state.js setter generator, fetchJson() helper, makeDraggable() extraction, encoder base-class close(), update_index/update_index_async unification, InvokeMetadataView strategy class) are intentionally left out of this batch — each is its own reviewable change.

Test plan

• ruff check photomap tests — clean
• npm run lint + npm run format:check — clean
• pytest tests/backend — 257 passed (same count as before)
• npm test — 288 passed across 18 suites
• Manual: /curate (async) and /curate_sync both return identical-shaped responses for the same request
• Manual: bookmark a search result via the seek-slider's star — confirm exactly one toggle (the prior double-bind would cancel itself)

🤖 Generated with Claude Code