feat(embedder): SQLite per-utterance embedding cache

[voorhs]

Summary

Replaces the .npy-file-per-call embedding cache with a single SQLite database keyed per utterance, and lifts the triplicated cache code out of the three backends into one template method on BaseEmbeddingBackend.

Why: the old cache keyed on hash(model + entire_utterance_list + prompt), so any change to a list (reorder/add/drop one item) was a full miss, and every distinct list became its own .npy file (inode growth, no atomicity, no concurrency story). Per-utterance keying stores each (model, utterance, prompt) once — overlapping calls now reuse the overlap.

Honest scoping: SQLite does not make warm cache hits faster (they were already sub-ms). The wins are correctness (atomic writes), operability (one file, concurrency, eviction groundwork), and enabling per-utterance keys without an inode explosion.

Scope (decided with maintainer up front)

• ✅ Embedding cache only — the structured-output / LLM cache (generation/_cache.py) is untouched.
• ✅ Per-utterance keying — one row per (model, utterance, prompt), float32 BLOB.
• ✅ Eviction = groundwork only — created_at / last_accessed / size_bytes / model_hash columns + indexes; cache stays unbounded by default (no behavior change).
• ✅ Fresh start — old .npy caches are left as orphans (true migration is infeasible; old keys are list-hashes). One-time recompute on first run.
• ✅ AUTOINTENT_CACHE_DIR env var to relocate the cache (defaults to the OS cache dir).

What changed

• New autointent/_cache_dir.py — get_cache_dir() (env var + appdirs fallback).
• New autointent/_wrappers/embedder/_sqlite_cache.py — SQLiteEmbeddingCache (WAL, busy_timeout, BEGIN IMMEDIATE schema init with post-lock version re-read, INSERT OR IGNORE, model_hash read filter, chunked IN (...), graceful degradation to recompute on any cache I/O error), utterance_key, get_embedding_cache.
• Refactor BaseEmbeddingBackend.embed → concrete template (split hits/misses, dedup, reassemble in order); backends implement _embed_uncached. HashingVectorizer opts out via supports_cache = False (its 262k-dim vectors would be ~1 MB BLOBs, and recompute is cheap).
• Removed embedder/utils.py (get_embeddings_path).
• Tests: pure-Python unit tests for the store + a global cache-dir isolation fixture; integration tests proving per-utterance reuse (DB has exactly 3 rows after ["x","y"] then ["y","z"]), dedup, order preservation, and empty-input behavior.

Process / provenance

This PR includes the design spec and implementation plan under docs/superpowers/ for review. Both went through adversarial review-to-convergence (spec: 3 rounds; plan: 1 round with empirical ruff/mypy validation), plus a final whole-diff review that caught and fixed one real bug (HashingVectorizer empty input raised StopIteration under sklearn ≥1.5 — now returns (0, n_features)).

Verification

• Local: ruff check . and mypy src/autointent tests (strict, py3.10) both green.
• Tests: per project convention, the full pytest suite is not run locally — it runs on CI for this PR. Please check the CI status here. The new cache unit tests are pure-Python (no model downloads); the integration tests use the pinned tiny ST model.

🤖 Generated with Claude Code