Summary

• On a cold vocab cache, concurrent /cluster_labels and /image_label requests both dispatch get_or_build_vocab_embeddings through asyncio.to_thread. Every caller saw cache_path missing, loaded the encoder, and re-encoded the full vocabulary before the first writer's atomic .tmp → rename landed — observed as 5+ duplicate "Building vocab embeddings cache at …" lines per cold start.
• Added a per-encoder threading.Lock registry around the build path in cluster_labels.py, with a re-check of _read_cached_vocab inside the lock so the second waiter picks up the first builder's atomic rename instead of redundantly re-encoding.
• Per-spec (rather than global) locks preserve the existing ability for two albums with different encoders to build in parallel; only redundant builds of the same encoder are serialized.

Test plan

• pytest tests/backend/test_cluster_labels.py — 38/38 pass, including a new test_concurrent_builds_are_serialized that spawns 4 threads against a gated fake encoder and asserts encode_calls == 1.
• ruff check photomap/backend/cluster_labels.py tests/backend/test_cluster_labels.py clean.
• Manual: open the app with a cleared ~/.cache/photomap/cluster_vocab/ and trigger several concurrent slide-drawer opens; log should show one "Building vocab embeddings cache" line followed by one "Vocab embeddings cached".

🤖 Generated with Claude Code