fix(scan): exclude per-scan extractedAt and warnings from enrichment stage hashes

[sricursion]

Fixes #347

Makes rescans of an unchanged schema hit the enrichment resume caches instead of re-sending every table and column description to the LLM.

Every connector stamps a fresh extractedAt on each introspection, and that timestamp (plus any per-run warnings) was folded into the descriptions/embeddings/relationships stage hashes via the whole-snapshot hash. A rescan therefore always produced a new inputHash, both persistent caches (the SQLite stage cache and enrichment-progress/descriptions.json) missed, and enrichment restarted from zero — the repeated ~4-minute description runs in the issue.

Approach

enrichment-state.ts now hashes only schema-identifying snapshot content: a shared stableSnapshotHashContent helper strips extractedAt (fresh wall-clock per scan) and warnings (per-run diagnostics, e.g. one transient failed sample would otherwise re-key every stage) before hashing, and all three stage-hash functions route through it. This mirrors what stableLiveDatabaseHashContent in local-stage-ingest.ts already does for the ingest work-unit cache. No enrichment stage reads either field, so the hash loses no discriminating content, and a real schema change (added/renamed column) still re-keys.

Existing caches re-key once after upgrading (the hashed content shape changed); rescans are stable from then on.

Tests

• New regression test: two snapshots differing only in extractedAt/warnings hash identically for all three stages, and an added column still re-keys all three. Fails on main, passes here.
• Before/after check against the built package: with dist from main, a fresh extractedAt re-keyed all three stages; with this change all three hashes are stable.
• type-check, CLI test suite, and dead-code (biome + knip default + production) pass. I developed on Windows, where ~60 test files fail on main for pre-existing environment reasons (sqlite EBUSY on temp-dir cleanup, doubled Z:\Z:\ drive prefix in benchmark fixture paths); I diffed failing-file sets between main and this branch on the same machine to confirm no new failures. Happy to file those Windows issues separately.

Out of scope

Per-table checkpointing so a scope/schema change only regenerates changed tables — tracked in #348.

[vercel]

@sricursion is attempting to deploy a commit to the Kaelio Team on Vercel.

A member of the Team first needs to authorize it.