feat: LanceDB semantic vectors + in-process native Qwen3 embedder (MinSync shape)

[vkehfdl1]

Fixes #39 (also supersedes the closed #38 — batching alone was not enough).

What changed

Storage: semantic_vectors (JSON number[] text in SQLite) is replaced by a LanceDB table at <data-dir>/semantic.lance — a typed FixedSizeList<Float32> vector column plus account/mailbox/thread/date predicate columns. SQLite remains the source of truth for messages, chunks, and FTS; the Lance table is a rebuildable semantic index (same layout MinSync produces: _versions, _transactions, data). <data-dir>/semantic.identity.json persists the embedder identity next to the table.

Embedder: the zero-config default is now the in-process native profile — native:Qwen/Qwen3-Embedding-0.6B (1024-d) through ONNX Runtime's native Node binding (WebGPU → Metal on Apple Silicon, automatic CPU fallback), last-token pooling, no HTTP sidecar. EmbeddingGemma stays as the legacy-onnx opt-in (and imports compatible legacy JSON vectors on first index); loopback-http remains the explicit #31 override.

Indexing: mandatory configurable batching (--batch-size / MAILCRAWL_EMBED_BATCH_SIZE, default 4 for the native profile, mirroring MinSync). Each batch commits the Lance upsert and the queue completion; a crash resumes from the last committed batch. Reuse is decided against stored content hashes so re-enqueued rows never re-embed.

Query: searchSemantic is a cosine vector-table query with account/mailbox/thread/from/after/before pushed down as Lance SQL predicates (only to hydrates through SQLite). Search against a table built by a different embedder fails loudly with a rebuild hint.

Rebuild guard: identity mismatch (provider/model/dimension/prefixes) on mailcrawl index discards the table and re-embeds everything — never silent reuse. repair --semantic forces a full rebuild; doctor/status report healthy / interrupted / never-completed / rebuild-required / corrupt / missing from the store identity plus the queue.

Verified on this machine (M4 Max, macOS arm64)

• Real Qwen3-Embedding-0.6B on WebGPU/Metal: load ~1s, batch-of-4 ~30ms, dim 1024, sequential calls stable, sensible cosine rankings (related 0.91 vs unrelated 0.42).
• Full CLI e2e with the default native path: sync → index (2 embedded) → status/doctor healthy → search:semantic / --mode hybrid ranked hits → mailbox filter → resume (reused: 2) → repair --semantic (rebuilt: true) → mismatch guard exits 1 with the rebuild hint.
• npx tsc --noEmit clean, npm run build clean, 100/100 vitest tests pass (new tests/issue-39.test.ts: 15 tests over the real LanceDB binding — store roundtrip, filters, dimension guard, crash-resume batching, rebuild-on-identity-change, legacy import/discard, sweep, summaries), npm run validate:skill valid.

Notes

• PR fix: batch semantic embedding to prevent OOM kill #38 was closed unmerged; its batching behavior is included here as part of the redesign.
• Version bumped to 0.2.0 (semantic storage format change).
• Docs updated: README, docs/architecture.md, docs/multilingual-installation.md, skills/mailcrawl/SKILL.md, THIRD_PARTY_NOTICES.md (Qwen3 Apache-2.0, LanceDB Apache-2.0).