Add mrplex verify — read-only integrity scrub

[usergenic]

Summary

Adds mrplex verify, a read-only integrity scrub over the store — mrplex's git fsck. In an append-only store the version chain is the source of truth and every other table (FTS, links, chunks, embedding backlog) is a derived index that can silently drift; verify re-derives what should be derivable, diffs it against what's stored, and reports inconsistencies as structured findings without ever writing. Implements docs/verify-plan.md in full (WS1–WS5).

kernel.verify(ctx, spec) returns a VerifyReport — findings are data, never exceptions, so a clean store returns an empty list and a corrupt one still returns a report. Six check families:

• chain — prev/next symmetry, one-current-per-doc, one-live-per-path, cycles, repo-id drift, orphan documents
• hash — recompute contentHash vs. the stored content_hash
• frontmatter — frontmatter_raw ↔ frontmatter round-trip, $-intrinsic leaks
• fts — SQLite fts_docs rowid ↔ version bijection (skipped-with-note on Postgres, whose fts_tsv is a generated column that can't drift)
• chunks — embedding provenance (nonexistent-version orphans, backlog orphans, mixed-dim)
• links — re-extractEdges vs. stored rows + resolution correctness

Findings never repair; each carries a suggested_fix pointing at the relevant backfill.

Surfaces

• CLI mrplex verify — --check (repeatable), --severity, --max-findings, --json, --ci (exit 1 on any finding at/above threshold). Repo optional: omit -r for a whole-store scan, required for the whole-store fts/chunks-orphan checks.
• MCP verify tool with full outputSchema (25 tools now).
• REST GET /verify (whole-store) and GET /repos/{repo}/verify (scoped); GET-only, no ETag.
• Threaded through KernelClient (local + remote) and the auth shell (read-only, scope-narrowing).

Notable design refinements found during implementation

• FTS is engine-asymmetric — SQLite membership is a whole-history rowid↔version bijection read via the fts_docs_docsize shadow table; Postgres has no separate FTS structure. The fts family is SQLite-only via an optional VerifyFtsScans capability.
• "Orphan" means a nonexistent version, not a superseded one — the embed worker deliberately keeps chunks on superseded versions.
• chunks / embedding_backlog / fts_docs are global tables — their orphan checks are whole-store, skipped-with-note under --repo.

Cost

verify is O(total versions) — it walks history, unlike everything else on the read path. Keyset-paginated throughout; --check/--repo narrow the work; findings capped with exact counts.

Test plan

• npm run typecheck clean
• Full suite green: 1144 passed, 2 Postgres suites skipped (no MRPLEX_TEST_POSTGRES_URL)
• Corruption-injection tests prove each finding fires (test/verify-checks.test.ts, 21 cases)
• Adapter parity + corruption scans (test/verify-scans.test.ts)
• CLI / MCP / REST surface tests (test/cli-verify.test.ts, http-mcp, http-rest)
• Postgres parity for the new adapter methods is written but unverified locally — needs a CI run with MRPLEX_TEST_POSTGRES_URL set

🤖 Generated with Claude Code