fix: Report clustering availability and publication health

[buildingjoshbetter]

Scope

Truthful clustering availability and maintenance health for #720, stacked on #808. Optional clustering packaging is unchanged. This adds observations and categorical warnings; health reads do not open a database, load a model or schedule maintenance.

Changes

• Add clustering to engine stats and MCP health/status while preserving the existing five health entries and their binary ok/degraded status contract.
• Report availability, last attempted outcome, successful output count, source freshness, coverage, pending caller commit and bounded error categories separately. Empty success is distinct from missing dependencies, pending work and a failed attempt.
• Inspect only top-level module availability, package metadata and bounded existing SQLite checkpoint/dependency state. Installed packages alone cannot certify usable native code or current vector coverage.
• Give installation guidance only when the optional HDBSCAN package is actually absent. Busy foreground/model ownership returns a deferred observation without waiting or triggering work.
• Emit a warning after an actual unavailable/failed clustering attempt, deduplicated per canonical database and failure transition. Health polling is silent. Warning text contains no exception message or source content.
• Retain prior process failure evidence after a successful borrowed savepoint. Caller rollback cannot make failed clustering look recovered; a later committed successful maintenance pass clears that evidence.
• Catch clustering inspection failures at the MCP boundary so the five existing health entries remain available. Failed owned inspection rollback is reported categorically and engine ownership is released.

Validation

• 24 focused stdlib/SQLite tests passed, including missing packages, metadata failures, disconnected/busy state, pending/empty/failed/current generations, unverified coverage, committed recovery, caller rollback and warning deduplication.
• The existing binary health-contract test body passed unchanged. Three original-production negative controls failed their intended assertions with zero setup errors.
• Both borrowed-transaction regression cases (zero and nonzero output) failed before the correction, then passed through outer rollback and later committed recovery.
• Integration-based candidate with actual sqlite-vec: 2,766 passed, 22 skipped, 3 deselected; 1,865 subtests passed. Log SHA256: 18e52b104aeff0e5145bb49e42b3012a699032a61df8795501d4e88e2de20df0.
• Exact isolated publication branch with actual sqlite-vec: 2,650 passed, 22 skipped, 3 deselected; 1,251 subtests passed. Its execution-log hash is recorded in the maintenance checkpoint.
• Configured Ruff, Python 3.10 syntax, whitespace, new-content privacy and commit-identity checks passed. Production Python and focused test bytes match the reviewed integration candidate. Hosted checks are pending.

Limits

There is no outer-commit callback. A prior process failure therefore remains conservatively visible until a subsequent committed successful maintenance pass. This does not certify arbitrary later caller commits, full vector-generation coverage or direct external database edits.

The existing bulk-ingestion result may still contain its original raw exception message; this draft does not claim to sanitize every pre-existing error payload. Models, rerankers, inference precision, dimensions and retrieval limits are unchanged. No latency, process-memory, MPS or thermal improvement is claimed. Only generated synthetic test data was used; no personal memories, transcripts or credentials. No main merge, release or issue closure.