[malibio]

The reader this issue describes (_seed.key / _seed.version through the CLI) was replaced by #3803: seeding stamps _seed.key, config_version and guidance_version on each seeded root, the runner reads them from the served database, fails an empty read-back, and aborts on drift between reps.

The remaining acceptance criterion was a test of the read-back against the shape the store holds. This adds one that runs the real sqlite3 against a table shaped like the daemon's, including an unstamped user skill and a non-guidance node that carries a stamp.

🤖 Generated with Claude Code

[malibio]

Code Review: PR #3804 (issue #3801)

Review Type: Initial review (no prior reviews or comments).

Summary

Test-only PR: one test in scripts/eval/preflight.test.ts runs the real sqlite3 against a node table shaped like the daemon's and checks readGuidanceProvenance. The issue body is stale: the _seed.key/_seed.version reader it describes was already replaced by #3803, and this PR correctly adds only the missing real-database coverage.

Independent verification of the stored shape

• packages/core/src/markdown/mod.rs prepare_nodes_from_template (~L1451-1467) stamps properties._seed on the root node with key (template title), config_version, guidance_version, tier. The comment at ~L1444-1450 confirms _-prefixed keys are exempt from namespace hoisting, so _seed sits at a fixed top-level path regardless of node type.
• packages/core/src/services/node_service/mod.rs reconciliation (~L2191-2240) reads the same _seed.config_version / _seed.guidance_version, so the writer and the reconciler agree with the eval reader.
• packages/core/src/db/schema.rs: node(id, node_type, properties TEXT, ...). The test's reduced table is compatible with the query readGuidanceProvenance runs (json_extract(properties, '$._seed.key'), etc.).
• Reader: scripts/eval/preflight.ts:527-581 reads _seed.key, config_version, guidance_version for agent-guidance and skill; throws on an empty type. Drift: checkGuidanceDrift (scripts/eval/runner.ts:815) fingerprints per-type, sorted key@config/guidance; runner calls it per rep (runner.ts:1594) and aborts, with unit tests in runner.test.ts:514.
• bun test scripts/eval/preflight.test.ts: 38 pass, 0 fail.

Requirements Check

• provenance.guidance lists every seeded guidance and skill node with a version: met by main (Eval preflight: read guidance provenance from the database, refuse to run beside another model (closes #3788) #3803). Every seeded node with _seed.key is listed with its config and guidance version; a type with none fails the run. This PR now proves it through real SQLite.
• A rep whose guidance differs from rep 1 aborts: met by main (checkGuidanceDrift, runner.ts:1594, tested in runner.test.ts).
• A test covers the read-back against the daemon's shape: met by this PR. The test covers the nested _seed shape, a user skill without a stamp (excluded), and a non-guidance type carrying a stamp (excluded).
• Issue wording "delete the _seed reader": satisfied differently. _seed is the correct source and the stale reader no longer exists.

Findings

• 🟢 scripts/eval/preflight.test.ts:115-146: Sound. SQL is built from literals only, the tmpdir is cleaned in finally, and the sqlite3 exit code is asserted so a missing binary fails loudly rather than silently.
• 🟢 Nit: the test's table has 3 columns versus the daemon's 9 (STRICT). Fine for what is exercised; if the reader ever filters on lifecycle_status or similar, the fixture must grow.
• 🟢 Nit: the test comment says seeded "by prepare_nodes_from_template"; that stamps the root only, and children carry no _seed. Accurate for the test, no change needed.
• 🟡 Process (non-blocking): issue Eval guidance provenance reads _seed keys that nothing writes #3801's body is out of date relative to main (it describes a reader that Eval preflight: read guidance provenance from the database, refuse to run beside another model (closes #3788) #3803 already removed). Suggest a closing comment noting Eval preflight: read guidance provenance from the database, refuse to run beside another model (closes #3788) #3803 delivered the reader and drift check and this PR the test, so the history reads correctly.

Recommendation

APPROVE (self-authored; posting as COMMENT). Net-positive, test-only, no production code touched. Merging is the user's call.

[malibio]

Review verdict: APPROVE (self-authored, so posted as a comment). Test-only; seeding stamps top-level properties._seed.{key,config_version,guidance_version} as verified in core; all three acceptance criteria are met by main (#3803) plus this PR. Full report in the comment above.

[malibio]

Address review

Review verdict: APPROVE. No code changes were needed.

• ⏭️ Skipped (🟢): fixture has 3 columns versus the daemon's 9. The reader selects only node_type and properties; revisit if it ever filters on lifecycle_status.
• ⏭️ Skipped (🟡): stale issue body. The PR's closes #3801 carries the explanation (the reader and drift check landed in Eval preflight: read guidance provenance from the database, refuse to run beside another model (closes #3788) #3803, the read-back test here).

Re-Review Decision

Decision: NO RE-REVIEW NEEDED

Rationale: no commits since the review; the PR is a single test addition.