fix: fail closed on unhealthy prewrite search

[VailElla]

Summary

• expose per-backend search health and return a non-zero status when the primary SQLite backend is unavailable
• bound long prewrite descriptions before search while retaining the full input for safety and similarity checks
• fail closed in prewrite and postwrite reconciliation instead of treating search failures as an empty result set
• improve long-description matching with reverse coverage and normalized title detection
• require sufficiently specific candidates before title or reverse-coverage shortcuts can force an update

Root cause

The closeout wrapper treated search-process failures as an empty candidate list. That made an unhealthy primary index indistinguishable from a legitimate no-match result, so prewrite could recommend ADD. Long descriptions also diluted one-way lexical coverage enough to miss an existing memory even when its title was present.

The matching shortcuts need their own guard: a short generic title or a candidate with only a few tokens can otherwise achieve high reverse coverage from an incidental mention. This revision limits those shortcuts while preserving specific English and Chinese title matches.

Impact

Prewrite now returns status=error and recommendation_unavailable_reason=reconcile_search_unhealthy when SQLite search is unavailable. Postwrite reconciliation records a blocking finding for the same condition. Ordinary failures from the optional vector backend remain non-blocking warnings when SQLite is healthy.

The branch is rebased onto 3cc6b43 and includes the updated postwrite test contract introduced on current main.

Safety and privacy

• the change contains only public scripts, documentation, and synthetic tests
• no real memory vault content, local configuration, database, logs, model data, credentials, or personal absolute paths are included
• both commits use the GitHub noreply author and committer address
• the full diff passed private-path, credential, private-key, forbidden-artifact, and whitespace scans

Validation

• /usr/bin/python3 -m unittest discover -s tests -p 'test_*.py' — 38 tests passed
• Python 3.14 targeted reconciliation and search tests — 13 tests passed
• isolated CLI test with a missing SQLite index — exits 2 with no write recommendation
• isolated CLI test with a healthy template index — returns UPDATE for the existing workflow
• /usr/bin/python3 -m compileall -q scripts tests
• /usr/bin/python3 scripts/agent_memory_check.py --skip-state-db
• git diff --check upstream/main