feat(score): add Rust profile (#[test] / tokio::test / rstest / proptest)

[hossein-webdev]

Adds a rust profile to scripts/score.py, following the existing per-language dict pattern, plus 13 named regression tests. Rust felt like a conspicuous gap in a benchmark about test quality — it has strong idioms for every axis you measure.

Calibrated against three real, well-tested suites in the same spirit as the kotlin/swift work: serde_json, rust-lang/regex, and RustCrypto/hashes.

Rust-specific findings

Each of these is pinned by a test, because each one silently wrecked the numbers before I caught it:

1. Inline #[cfg(test)] mod tests is Rust's dominant unit-test idiom. A tests/-only file pattern scores those suites as empty, so test_file accepts .rs generally — the same permissiveness the python profile already has.
2. #[cfg(test)] must not count as a test definition, or every inline test module inflates test_count by one and flatters the D.1/D.2 denominators. The pattern requires test (or ns::test) immediately after #[, which cfg( doesn't satisfy.
3. rustfmt splits long asserts across lines — and long literals are exactly the ones that get split. A line-anchored B.1 scored serde_json's suite at 12 against a hand count of 36. The expected-last branch now scans [^;]{0,200}?, which crosses newlines but can't cross a statement boundary. (Same shape as the note in your go B.1 about table-driven suites scoring ~0.)
4. Raw strings r#"…"# are the embedded-JSON fixture idiom and legitimately contain ", so a [^"\n] class stops at the first inner quote and scores serde_json's JSON vectors at zero. They get their own branch.
5. Numeric crates park hex vectors inside tuples and slices — assert_eq!(big.hi64(), (0xA000000000000000, false)), from_u32(&[0x00140000, 0x140000]) — so anchoring hex to the comma missed them.

Two deliberate design calls, both matching existing precedent:

• A.2 is None (uncountable), as in Go and Swift. A #[cfg(test)] module sees private items by design, and integration tests under tests/ can only reach pub — there's no countable private-access smell, and returning 0 would hand the axis a free win.
• rstest's bare #[case] — the marker for which function argument receives the row — is excluded from param; only #[case(...)] rows count. Counting both gave every parametrized test one extra. (My own test caught this.)

Accuracy

Heuristic, same tier as the other non-Python profiles. Measured against a hand count on the calibration corpus:

suite	scorer B.1	hand count
rust-lang/regex	3	3
serde_json	41	36

serde_json over-counts by 5 — the bounded scan occasionally picks up a hex literal from an adjacent statement. Trustworthy for trends and worst-offenders, as your docs frame the non-validated profiles; "validated": False is set accordingly.

Testing

python -m pytest tests/ -q → 85 passed, 4 xfailed (the 4 xfails are the pre-existing ones, untouched).

Counts in the new tests are hand-derived by reading each fixture, not recomputed with the profile's own regex.

Also updated

--lang lists in SKILL.md and README.md, and the module docstring's supported-language block and calibration note.

---

Unrelated, and happy to send it separately if useful: render() emits a ↓ that raises UnicodeEncodeError on Windows consoles using cp1252 (--json is unaffected). Repro: python score.py --tests <dir> on Windows without PYTHONIOENCODING=utf-8.

Thanks for publishing this — the rubric is the most useful thing I've read on test quality in a while, and it's now the basis of a test-quality skill in an MIT skill pack I maintain, credited to you in the README and the skill itself.