test(gq): add the GQ logic test harness, corpus, and fix-regression gate

[azimafroozeh]

What & why

Implements RFC 0045 (GQ logic tests, #584): a query-level regression becomes one small text file instead of a Rust test, and issue-closing PRs are held to carrying one.

• crates/omnigraph/tests/gq_logic_tests.rs walks tests/gq_logic_tests/*.gqt: one self-contained case per file (issue anchor and red_on: provenance, then schema, seed, and query / mutate / restart steps with per-step params and expect).
• Each case runs on the real engine: fresh store, init, load, index builds only when the case needs them, steps in order under a pinned traversal mode.
• Expect modes: unordered (multiset), ordered, error: <substring>, affected: nodes=<N> edges=<M>; number normalization never routes integers through f64.
• The parser is fail-closed: every malformed input is a refusal, and every refusal has a self-test.
• OMNIGRAPH_GQ_BLESS=1 rewrites a failing case's expect rows in place; OMNIGRAPH_GQ_LOGIC_TESTS=<substring> filters cases by file name.
• First corpus, five cases: the two performance: bm25/rrf ranked read + a join materializes the joined column corpus-wide → Arrow 2 GB offset overflow at ~700k+ rows #563 regressions, an ordering-refusal pin, a foreach + mutation + restart round-trip, and a typed mutation error.
• Two new required PR checks in gq-logic-tests.yml: GQ Logic Tests runs the target pre-merge (the workspace job keeps covering it post-merge), and Fix Regression Gate (scripts/check-fix-regression.py) holds every issue the body closes by keyword to a matching issue_N test or .gqt case in the diff, waivable per PR with the no-repro label.
• AGENTS.md gains the contract sentences: logic tests are the default query-behavior medium, every issue fix lands a regression at the cheapest tier that catches it, and #[ignore] messages open with their species.

The nightly heavy-repro job stays deferred until its first tests/repro_issue_*.rs member lands (the job fails on an empty glob by design).

Backing issue / RFC

• Implements RFC 0045, docs/rfcs/0045-gq-logic-tests.md (docs(rfc): add RFC 0043, GQ logic tests #584). docs(rfc): add RFC 0043, GQ logic tests #584 merges first; its implementation: flip rides that branch.

Checklist

• Change is focused (one test target, its corpus, its two CI checks, the AGENTS.md contract, dev-doc edits; no src/ changes)
• Tests added/updated for behavior changes (84 harness tests: 5 corpus cases plus one self-test per refusal; the gate script carries a self-test the CI job runs first)
• Public docs updated if user-facing surface changed (dev-facing only: testing.md, ci.md, branch-protection.md, AGENTS.md)
• Reviewed against docs/dev/invariants.md (test-and-CI diff; no invariant or deny-list item touched)

Local verification

• cargo test -p omnigraph-engine --test gq_logic_tests — 84/84
• cargo clippy -p omnigraph-engine --all-targets -- -D warnings -W clippy::dbg_macro, cargo fmt --all — clean
• scripts/check-fix-regression.py --self-test — ok; plus three scenario runs against this branch's own diff: closing performance: bm25/rrf ranked read + a join materializes the joined column corpus-wide → Arrow 2 GB offset overflow at ~700k+ rows #563 passes (the corpus matches), closing an issue with no test exits 1, the no-repro label waives
• scripts/check-docs.py, scripts/check-agents-md.sh, scripts/check-workflow-action-pins.py — all OK
• cargo test --workspace — not run locally (no src/ changes); CI covers it

Notes for reviewers

• The only engine seam used is the test-only with_traversal_mode pin.
• Two maintainer actions after merge: create the no-repro label, then run scripts/apply-branch-protection.sh (running it earlier leaves open PRs pending on contexts they cannot receive; both new checks already report on this PR).
• The gate reads GitHub's keyword forms fixes #N, fixes: #N, and fixes:#N (leading zeros normalized); URL, cross-repo, GH-N, and no-space fixes#N closings pass unexamined and stay review's job.
• The edited/labeled/unlabeled PR types re-run the gate when a body edit or the label changes its answer; the test job re-runs on them too (accepted cost).
• The PR cache restores but never saves (save-if: push), so per-PR caches cannot evict main's.
• Stricter than the RFC on small points (traversal pin forces index builds, 10000-iteration loop cap, exact-spelling numeric tokens, CRLF and foreign corpus files refused); each flagged on docs(rfc): add RFC 0043, GQ logic tests #584.

Greptile Summary

The PR adds a query-level logic-test harness and corpus, then integrates it with pull-request regression enforcement.

• Adds the .gqt parser, runner, result normalization, blessing workflow, and initial regression cases.
• Adds required GQ test and fix-regression CI checks.
• Documents the new testing and branch-protection contracts.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
scripts/check-fix-regression.py	Tightens regression detection to executable-looking .gqt cases and Rust test definitions; the previously reported broad text-match bypass has been addressed within the script's documented boundary.
crates/omnigraph/tests/gq_logic_tests.rs	Adds the parser, execution harness, comparison modes, restart handling, filtering, blessing support, and extensive refusal self-tests.
.github/workflows/gq-logic-tests.yml	Runs the logic-test corpus and regression gate on the relevant pull-request events.
.github/branch-protection.json	Adds the two new workflow contexts to the required checks.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  PR[Pull request] --> Tests[GQ Logic Tests]
  PR --> Gate[Fix Regression Gate]
  Corpus[Changed .gqt and Rust tests] --> Gate
  Body[Closing keywords in PR body] --> Gate
  Tests --> Protection[Required branch checks]
  Gate --> Protection

Loading

_{Reviews (2): Last reviewed commit: "test(gq): tighten the fix-regression gat..." | Re-trigger Greptile}

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.