CSS at-rules (@layer / @media / @supports) counted as functions; css tree-sitter-accuracy baseline stale (real_functions 24→25)

[squid-protocol]

Summary

tree-sitter-accuracy-audit fails on css against the pinned v1.2.0 corpus:

=== css ===
tree_sitter_accuracy_audit: ground-truth metric(s) drifted from the pinned corpus's expected values:
  real_functions: baseline=24 current=25
This means the corpus changed. Investigate before regenerating.

This is pre-existing on main (HEAD 4cc5001f): PR #2500 ("css tri-comparison … audit-reader fixes, no engine change") changed the audit walk but did not re-bless tests/tree_sitter_accuracy_baseline_css.json, which was last regenerated by #2478's bulk v1.2.0 pin bump. Surfaced by the CI run on PR #2506 (an unrelated COBOL prism fix); filing separately per scope discipline. Also noted as a follow-up in the v1.2.0 release notes.

Two distinct things going on

1. Stale baseline — GitGalaxy actually improved

python tests/tools/tree_sitter_accuracy_audit.py --lang css --regenerate on current main:

                        committed   current
found_functions              20         25
extra_functions               5         0     <-- GitGalaxy went from 5 FP extras to 0
found_classes               227        227
real_functions               24         25
args_comparable              20         25
args_exact_match             14         19

The extra_functions: 5 → 0 is a real gain already sitting on main (from #2500's audit-reader fixes) that the committed baseline doesn't reflect.

2. Real false-positive — CSS at-rules parsed as functions-with-args

The audit's args comparison flags GitGalaxy counting CSS at-rules as function definitions:

File	Construct	GitGalaxy
gutenberg_css_modules/button.module.css:3	@layer wp-ui-components { … }	layer — 1 arg
tailwindcss_atrules/preflight.css:291	@supports (not (-webkit-appearance: …)) or (contain-intrinsic-size: 1px) { … }	supports — 3 args
threejs_app_ui/editor_main.css:660	@media all and ( max-width: 600px ) { … }	media — 3 args
threejs_app_ui/editor_main.css:720	@media …	media — 4 args

real=0 in every case — tree-sitter does not treat these as functions. This matches the known "css precision 100%→83.3% (@layer/@media at-rules counted as functions)" regression flagged at the v1.2.0 cut.

Proposed work

1. Tighten the css func_start rule in gitgalaxy/standards/language_standards.py so @layer / @media / @supports / @container / @scope (and other prelude-taking at-rules) are not anchored as function definitions. A real CSS function is @function name(...) (very new) or a custom-property/calc()-style call — not a conditional-group at-rule.
2. Investigate the real_functions 24 → 25 ground-truth bump (which corpus file / which node tree-sitter now counts) before regenerating — the audit message explicitly asks for this.
3. Once (1) and (2) are settled, regenerate tests/tree_sitter_accuracy_baseline_css.json (+ the summary table in language_standards.py) and re-bless the golden masters if css output shifts.

Repro

python tests/tools/tree_sitter_accuracy_audit.py --lang css --ci      # fails
python tests/tools/tree_sitter_accuracy_audit.py --lang css           # shows the arg mismatches

🤖 Generated with Claude Code

[squid-protocol]

PR #2509 addresses this.

Item 2 (investigate the real_functions 24→25 bump): it is not corpus drift. files_scanned is 21 in both the committed baseline and now — #2478 already regenerated this baseline against the v1.2.0 pin. The bump is #2500's own _get_node_name audit-reader changes (keyframes branch un-deadcoded → +funcs; bodyless @layer a,b; now None → −funcs; net +1) landing without re-blessing tests/tree_sitter_accuracy_baseline_css.json. #2500 updated the tri-comparison chart/ledger and docs/language_status/css.md but missed this file; its commit message records the target state ("css Func Found 25*/24*/0* -> 25/25/0"). Zero engine change, zero corpus change.

Item 3 (regenerate baseline + summary table): done. css → 100% func recall / 100% func precision, matching the tri-comparison. No golden-master re-bless — no engine/regex change.

Item 1 (tighten css func_start so at-rules aren't anchored as functions): declining, and I think the issue's framing here is superseded by #2500. The three css function entries in tri_comparison_ledger.json are all status: validated, with css/function/existence/agree[gitgalaxy,tree_sitter]_vs[ctags] (25 occ) explicitly ruling GitGalaxy correct: at-rule keywords are "the closest function-shaped construct CSS has." tree_sitter_accuracy_audit.py's NODE_MAPS also deliberately counts media_statement/supports_statement/keyframes_statement/at_rule as ground-truth functions (#1313). Removing them from func_start would revert a validated design decision and crater measured recall. The "precision 100%→83.3%" regression flagged at the v1.2.0 cut was already resolved by #2500 as an audit-reader bug, not an engine defect.

The one true residue — GitGalaxy reading a nonzero args count off an at-rule prelude (@media all and (max-width: 600px) → 3) while tree-sitter reports real=0 — is an informational args_comparable gap (the gated args_exact_match is unaffected and improved 14→19). PR #2509 documents it in the audit's SCOPE docstring rather than chasing it, same "no formal signature at the declaration site" shape as the shell/perl note (#1518/#1519).

Proposing to close this once #2509 merges.