fix(chroma): flag stale HNSW ghost surplus

[fallenmi]

What does this PR do?

Refs #2176.

After repeated or bulk deletes, Chroma can retain substantially more persisted HNSW elements than live SQLite embeddings. The reported case had 1,606 SQLite rows and 9,974 HNSW elements (divergence=-8,368), but the capacity probe classified every negative divergence as OK and explicitly called the surplus safe.

This PR fixes the diagnostic and recovery path:

• large HNSW surpluses now return status="stale" and repair_recommended=true;
• the existing diverged flag remains false for this direction, so the native-crash fence reserved for an undersized HNSW index does not disable vector reads or writes;
• surplus must strictly exceed the same configured threshold used by the capacity probe; small persisted delete lag remains OK;
• repair-status prints STALE and recommends rebuilding from SQLite, including when all live drawer rows were deleted but ghost HNSW elements remain;
• MCP emits one warning per distinct stale verdict instead of reporting the state as within tolerance or spamming every refresh;
• positive SQLite-over-HNSW divergence now also exposes the additive repair_recommended field.

Scope

This is deliberately a detection and recovery-guidance fix. It does not compact or prevent Chroma's deleted HNSW elements, and it does not add cumulative-delete accounting or a hint to every delete_by_source response. Those can be handled separately without coupling a potentially expensive repair to the write path.

Safety

diverged=true currently activates the #1222 fail-closed path: BM25-only search, vector-write refusal, and several preflight aborts. The surplus reported in #2176 degraded ranking but was not shown to create the undersized-index native-crash risk, so the new stale state recommends repair while intentionally leaving vectors enabled.

How to test

python -m pytest -q \
  tests/test_hnsw_capacity.py \
  tests/test_repair.py \
  tests/test_mcp_server.py

pre-commit run --all-files

Local results on current develop:

• 552 targeted consumer tests passed;
• 4,301 repository tests passed, 31 skipped, with only the unchanged environment-sensitive test_init_filters_sys_path_from_leaked_pythonpath subprocess assertion excluded;
• Ruff check, Ruff format, pre-commit, and git diff --check passed.

Regression coverage pins the exact 1,606/9,974 report, explicit and legacy threshold boundaries, zero-live-row ghost state, repair output, vector-fence behavior, and warning deduplication.