Add a revision-safe persistent embedding cache

[dajiaohuang]

Summary

• add an optional process-safe SQLite embedding cache behind each domain prefix
• cap it by entry count and logical bytes, validate float32 payloads, and fail open on lock/corruption
• pin built-in BGE profiles and require custom models to use immutable 40-character commit SHAs
• scope SQL rows, memory keys, Chroma collections, and vector outbox jobs by model revision
• prevent pending vector jobs from being delivered or completed under another revision
• coordinate first-use initialization across local embedder instances and restrict new POSIX cache paths
• document privacy, tenant-isolation, capacity, and operational settings

Safety properties

• raw source text is not stored in SQLite (only model/text digests and vectors)
• low-entropy digest and embedding sensitivity are explicitly documented
• cache failures never block model inference
• cache use is opt-in and should stay outside source checkouts/public backups

Validation

• uv run --extra dev pytest (308 passed, 14 skipped)
• uv run --extra dev pytest tests/test_embeddings.py tests/test_vector_jobs.py -q
• uv run --extra dev ruff check .
• git diff --check
• verified the three built-in revisions against their official Hugging Face repository HEADs

Configuration

Set _EMBEDDING_CACHE_DIR; optional limits are _EMBEDDING_CACHE_MAX_ENTRIES, _MAX_BYTES, _BUSY_TIMEOUT_MS, and _EPOCH.