fix(knowledge): deflake projection integration tests by ranking vector assertions exactly

[waleedlatif1]

Summary

• knowledge-projection and unfilled-projection-source integration tests intermittently got [] from the vector leg for chunks the access predicate admits
• Cause: the vector assertions used permitted: { kind: 'unbounded', broad: true }, which walks the shared HNSW graph. The files sharing the CI database fill it with degenerate vectors ([1, 0, …], constant rows), so a fixture chunk can be pruned from every neighbour list and becomes unreachable however far the iterative scan goes. The one-hot fixture vector from fix(cli): strip endpoint trailing slashes in linear time, deflake projection test #8285 avoided ties, but not this
• Switch the vector assertions to broad: false. With no per-source index, this ranks the sliced sources exactly, under the same on-row visibility predicate the walk applies, so what the tests assert (revoke/grant/unfilled visibility) no longer depends on ANN recall

Type of Change

• Bug fix

Testing

• Reproduced the unreachable-row behaviour on pgvector 0.8.6: with a crowd of duplicate [1, 0, …] rows, an iterative walk with a 100k tuple cap returned 3 of 5 rows sitting at distance 0
• Ran the full CI shard (25 files, 255 tests) locally 5x on the fix: all green
• bun run lint, check:audits, docs-manifest:check pass

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
docs	Skipped		Sep 25, 2026 3:46am UTC

[cubic-dev-ai]

No issues found across 2 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

_{Re-trigger cubic}

[greptile-apps]

Confidence Score: 4/5

[Medium risk] Changes vector search ranking behavior in integration tests.

The PR appears safe to merge, with a non-blocking gap in broad vector-path integration coverage.

Findings

1.  Broad-path coverage is lost ▶

Summary

The PR changes vector assertions in two projection integration suites from a shared approximate graph walk to exact per-source ranking, avoiding flaky recall-dependent results.

• The visibility assertions remain in place for grants, revocations, and unfilled projection rows.
• The change leaves the broad graph-walk path without database-backed integration assertions.

_{Reviews (1) · Last reviewed commit: "fix(knowledge): deflake projection integ..."}