test(span-anchor): count #[attr] lines and trailing-comment return types as part of the declaration

[squid-protocol]

Summary

Follow-up to #3547. This fix was pushed to that branch after it had been squash-merged, so it never reached main.

The first CI run of the span-anchor gate listed sample units, and they showed two false positives in span_anchor_audit.py's rule:

• PHP 8 attributes: #[\Override] lines were read as # hash comments. So #[\Override] public function groupBy(...) counted as mis-anchored (laravel Collection.php).
• C return types with a trailing comment: int /* Return 1 on mismatch, else 0 */ on the line above cob_read_dict(...) was read as ending a construct, because the line ends in */ (gnucobol fileio.c).

Fix:

• A line starting with #[ is an attribute, never a comment.
• Trailing /* … */ and // … comments are stripped before the "ends a previous construct" test. That test now looks only at {, }, , and ;. Pure comment lines are still filler.

Baseline regenerated with span_anchor_audit.py --regenerate: PHP 22.1% → 21.9%, C 0.1% → 0%. Every other language is unchanged: TypeScript 18.1%, JavaScript 0.4%, the rest 0. Both shapes are pinned in test_span_anchor_audit.py.

Type of change

• Docs, tooling, or CI only

CI checklist

• pytest tests/tools/test_span_anchor_audit.py: 4 passed, run from /tmp via python -m pytest.
• ruff format / ruff audit clean.

Verification

$ python tests/tools/span_anchor_audit.py --regenerate --samples 3   -- 50 languages; php 21.9%, typescript 18.1%, javascript 0.4%, others 0
$ python -m pytest tests/tools/test_span_anchor_audit.py            -- 4 passed

graph-accuracy-audit.yml runs on this PR, because it touches the tool and the baseline.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YAtpBfGGyBGGovitSVaM5Q

---

Generated by Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab