Two defects surfaced by an adversarial audit of main after PRs #32–#35 merged. Each is a follow-up to code those PRs introduced.

1. Startup can silently orphan an existing KB registry (blocking)

PR #33 made the session/KB-metadata SQLite path resolve from config.database.path. Before that, every entry point hardcoded ./data/perspicacite.db and ignored the config key. The shipped config.example.yml — the documented starting point (cp config.example.yml config.yml) — sets database.path to ~/.local/share/perspicacite/memory.db. So a deployment created before #33 keeps its knowledge bases in the legacy working-directory file, while its config names a home-directory path that does not exist. After #33, honouring the config opens an empty database and list_kbs() returns an empty list: every knowledge base vanishes from the UI, and new metadata is written where the vectors do not follow.

#33's guard only defended the case where database.path is unset; the documented config sets it explicitly, so the guard did not fire for exactly the mainstream upgrade path.

guard_session_db_path now refuses to start when the configured database is absent and the legacy ./data/perspicacite.db still holds a non-empty kb_metadata table. The error names both paths and the fix. No data is moved — relocating a production registry implicitly is riskier than a loud stop. PERSPICACITE_ALLOW_NEW_DB=1 opts into a deliberately fresh instance. Wired into the web app, the MCP server and the two CLI entry points.

Verified against the real production registry and config: the guard raises with the correct message and paths, and the escape hatch bypasses it.

2. The embedding-health probe missed tail poisoning (advisory)

PR #32 added embedding_health to GET /api/kb/{name}/stats, sampling a bounded prefix (first 128 chunks). A mid-ingest embedding outage leaves a contiguous run of zero vectors, usually at the end of a collection, which a prefix-only sample reports as healthy. The probe now reads both ends, and the block gained total_chunks and complete so a clean sample of a large collection is not mistaken for a proven-clean collection.

Tests

• test_config_database_path.py: guard raises on the orphaning scenario; the escape hatch, an absent legacy DB, an empty legacy DB, and a config pinned at the legacy path all pass through.
• test_embedding_health_probe.py: zeros beyond the first 128 chunks flip degraded to true; a clean large sample is reported complete: false.

CI's unit suite: 2433 passed, 5 skipped, 0 failed. ruff clean on the new module; mypy clean on config/paths.py.