fix(cluster-labels): content-based vocab cache invalidation (survive reinstall)

[lstein]

Problem

The encoded tagging vocabulary was being rebuilt from scratch on first startup after a reinstall, even though the cached embeddings were still on disk and valid.

Root cause: both the vocab-embedding cache (_read_cached_vocab) and the per-album labels cache (_read_cached_labels) invalidated by comparing the cache file's mtime against the source vocab files. A pip reinstall rewrites the bundled cluster_vocab.txt with a fresh mtime even when its contents are byte-for-byte identical, so the mtime gate tripped and forced a multi-second CLIP re-encode of the entire vocabulary — despite the content being unchanged.

(The build was triggered legitimately, by a surviving per-device autotaggingEnabled preference or a second browser tab — that part is working as designed. This PR fixes the wasteful rebuild.)

Change

Switch both caches to purely content-based invalidation:

• _read_cached_vocab — drop the mtime gate. The exact phrase set is already stamped into the .npz and compared (set(phrases) != set(current_phrases)), which catches every real change (edits, additions, user-extras deletion/rename) regardless of mtime direction.
• _read_cached_labels — drop the vocab-mtime gate; rely on the stored vocab_fingerprint (sha256 content hash) already compared just below. The umap.npz and source-embeddings mtime checks stay, since those track album-local regeneration, not reinstalls.

The per-image in-process LRU (compute_image_label) intentionally keeps its mtime key: it's process-scoped, so a reinstall never reuses it anyway, and mtime is cheaper than hashing the vocabulary on that hot path.

Nothing here deletes the cache when autotagging is turned off — once encoded, the vocabulary stays cached across reinstalls for later use.

Tests

• Repurposed the two tests that asserted "pure mtime bump → rebuild" to assert "content edit → rebuild" (the real intent).
• Added regression tests proving a pure touch with identical content does not rebuild either the vocab-embedding cache or the per-album labels cache.

All green locally: 384 backend + 356 frontend tests pass, make lint clean.

🤖 Generated with Claude Code