test(gq): add the Kuzu-derived .gqt case corpus

[azimafroozeh]

What & why

97 kuzu_*.gqt feature cases (275 steps) translated from the Kuzu e2e tests (test/test_files/, MIT): regression coverage for match patterns, where, not { }, bounded traversal, return, order, limit, and the aggregates. Follow-up to #607. No harness change.

• Coverage: every statement in Kuzu's GQ-shaped areas that GQ can express. Of 1,690 statements, 1,132 have no GQ spelling (arithmetic, WITH, path variables, functions, multi-label patterns) and 278 were skipped (large datasets, Kuzu -SKIP); they are absent, not red.
• Where the engines disagree by design, the case pins GQ's reading and the header names Kuzu's value: {m,n} is shortest-path distance per endpoint pair, an undirected self-loop is one row, AVG at scale 12.
• 10 cases pin behavior the engine does not have yet and are red on this branch (listed below); the PR is a draft until they go green.
• crates/omnigraph-gqt/README.md gains the derived-case rules and the MIT notice; RFC 0045's Decision log records the source-prefix convention. The fix-regression gate keys on the issue_ prefix and never reads these files.

Backing issue / RFC

• RFC 0045, GQ logic tests (docs/rfcs/0045-gq-logic-tests.md, draft): Decision log amended with the derived-case rules; nothing here differs from its design.

Checklist

• Change is focused (one derived corpus, two doc paragraphs)
• Tests added/updated for behavior changes (the PR is tests; each case is its own test)
• Public docs updated if user-facing surface changed (none; crate README and RFC log updated)
• Reviewed against docs/dev/invariants.md — data-only diff, no invariant touched

Local verification

• cargo test -p omnigraph-gqt: 94 unit tests green; corpus 96 passed / 10 failed (the red-on-purpose set); slowest case 0.72 s.
• Every expectation checked against three oracles independent of the engine and of Kuzu: hand derivation from the seed with the Cypher re-read for fidelity; a SQL translation run on SQLite (268 of 270 steps agree, the 2 left are an oracle gap in delete-cascade counting; seeded faults: 86 of 86 values and 6 of 6 orderings caught); a Python evaluator of the documented semantics (234 of 247 agree, the rest being the hop-range cases where it let a path return to its start, which the self-loop fix's release note rules out). The pass found and fixed one wrong expectation, four limit steps that were green only by scan order, and five dropped reads.
• python3 scripts/check-docs.py: ok. Vocabulary guard, fmt, clippy: not run, no Rust or guarded surface in the diff.

Notes for reviewers

• Red on purpose, 10 files, each with a # notes: Red on purpose: line naming what it waits on. GQ Logic Tests is a required check, so this stays a draft; each fix PR carries its own minimal issue_N_ case and these remain as full-seed coverage:
  • sum over an integer column accumulates in f64 (1 file)
  • duplicate result column name, T25 refusal not yet in the compiler (2 files)
  • F32 cells render as the f64 widening, 0.99 as 0.990000009537 (4 files)
  • min/max over Bool and Date refused, T8 (2 files)
  • count($e) over an edge binding refused, T23 (1 file)
• 38 files inline the same seed: RFC 0045 keeps seeds inline by design; they are generated from one source, so a seed fix is one edit and a regenerate.

[chatgpt-codex-connector]

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