Auto-generated by engram codedb-insights. engram learns to re-rank codedb's results; this is where its learned policy disagrees most with codedb's lexical order, using git history as ground truth (commit subject → keyword = query, changed file = gold).

Problem

Over 30 git-derived queries (29 comparable), codedb's lexical search ranks the actually-changed file poorly: codedb MRR 0.356, vs 0.583 when re-ranked by engram's learned policy (centrality + eponymy). engram ranked the gold strictly higher on 19/29 comparable queries. 10/30 changed files weren't lexical hits at all — codedb's lexical ranker is structurally blind to them.

Evidence (codedb rank → engram rank · query → changed file)

Failing test case (red)

Per this repo's CONTRIBUTING.md ("Red-To-Green"), a failing case may be "an exact benchmark command and failing output". Reproduce from a codedb-indexed git checkout of this repo:

engram codedb-report .

Red — worst gap. Query correctness (gold = src/explore.zig, the file changed in that commit):

• Expected: the gold file ranks at Bug: edit.zig advisory lock not released on error paths #1.
• Actual: codedb ranks it at Bloom-augmented trigram postings (nextMask + locMask) #17, below lexical-only hits; engram's structure-aware re-ranker recovers it to Bug: edit.zig advisory lock not released on error paths #1.

Manual check: codedb search "correctness" should surface src/explore.zig near the top — today it is buried at #17.

Every row in the Evidence table above is an additional failing assertion (codedb rank > engram rank). Labels are git-derived (commit subject → query, changed file → gold): noisy, but the ranks are reproducible against a fixed index.