fix: Route style maintenance through coordinated ownership

[buildingjoshbetter]

Change

Style maintenance could bypass coalesced ownership or report current coverage without proving readiness in the checkpoint snapshot. This checkpoint routes startup, append notifications, explicit consolidation and bulk style work through the existing coordinator, with bounded metadata health and truthful coverage reporting.

Stacked on #814 (fix/l0-style-incremental-publish-747); checkpoint C for #747.

• Preserve eight public maintenance layers and nine consolidation result keys. Report private style status separately and retain sibling failure precedence.
• Keep one active worker and one pending request. Preserve the 25-insert retry threshold, cancellation, writer ownership and borrowed-transaction rollback behavior.
• Validate source tracking, hash/format metadata, owned output triggers and unsupported TEMP shadows in the same checkpoint snapshot. Bounded foreground health reads neither source/profile contents nor native models.
• Preserve all 274 original fixture assertions. Tests dedicated to public consolidation or vector rebuilding isolate unrelated style scheduling.

Validation

• Independently reviewed author v5: 2,946 tests and 2,344 subtests passed, 17 skipped, 3 deselected. Log SHA256: ff20d57a146b8091f26351d1b07f406921a4ef9f6bff265ccc6b959956dd6042.
• Corrected publication head 8cd28ac366f45e2642eabd8c26f3b8ac2ebc4785: 2,835 tests and 1,690 subtests passed, 17 skipped, 3 deselected. Unprivileged isolated Linux run with actual sqlite-vec, exit 0, no timeout or OOM. Log SHA256: 0870b605e6146d7f6f7ccb9932d1d91198ef57b85a6dc4838bf21760d779d90a. The original publication pass is retained with log SHA256 54578336462253d86dbe1def12e043cb9cff55c73aa88788e895199b1358cf5c.
• The original eight code/test files equal reviewed author 16ce28c571d04a2d7110d4655674592bf141baea. Documentation retains the exact prerequisite prefix plus the author's style-routing section. Remote source hashes matched the publication worktree.
• Independent publication review found no blockers. Ruff, Python 3.10 grammar and whitespace checks passed. Staged privacy scan: 2,120 added lines, nine files, zero findings.
• Earlier v1-v4 candidates and failed receipts are preserved. The v4 negative control fails eight assertions with zero errors; passing full suites did not replace adversarial readiness checks.

Fixture follow-up

Hosted Python 3.14 validation found a sparse surprise-score test expecting 1.45 but observing 1.05. Its shared fixture opened the engine with maintenance enabled, allowing a worker to replace the scores owned by the test. The follow-up disables unrelated maintenance before opening this fixture and closes its engine after each test, draining any unexpected coordinator. All 32 original assertions remain unchanged; production code is unchanged by this follow-up.

An independently reviewed, pinned synthetic probe gates the real coordinator worker around the foreground read. The original fixture reproduces the 1.05 score with one actual predictive rebuild; the corrected fixture creates no worker, retains the six seeded rows and returns 1.45. Both probes verify cleanup and no native model/network/subprocess calls. This demonstrates the failure mechanism; the historical hosted worker/interleaving remains unproven. Probe logs: original abf85eebc780e98487af2a7cd8dea37b3b2818d94746a322f98c2dd3d4301fe0, corrected 4b08ed44017d749f924e33a4a767f64669bdec43fcb84c75be7aba4990fee1fb. The failed hosted log and earlier probe portability failure are retained. Follow-up privacy and public commit identity checks found zero findings.

Limits

All workload data is synthetic. Models, rerankers, dtype, vector dimensions, retrieval depth and ordering are unchanged. Background quick_check may scan the database. Current trigger definitions cannot prove uninterrupted historical source tracking.

Native numerical equivalence, whole-process memory, Mac/MPS latency and thermal acceptance remain open. This draft does not close #747 or prove the original memory incident fixed. No main merge or release.