[squid-protocol]

Summary

Extends #1264's per-language class_start named-entity extraction to Zig, per epic #1295's checklist. Zig is the biggest/messiest language in this exercise so far (32 corpus files, real_classes in the hundreds), and the first to land with extra_classes/missing_classes still non-zero by design after investigation -- following the same discipline csharp's enum precedent in #1264 established (document real scope mismatches, don't chase every last case into one PR).

1. tests/tools/tree_sitter_accuracy_audit.py: ground-truth gap -- zig's error sets (const Foo = error{...};) get their own ErrorSetDecl node type, distinct from ContainerDecl, which NODE_MAPS["zig"] didn't include. GitGalaxy's own class_start regex intentionally matches error sets (its own comment says "Defines structural entities: struct, enum, union, error, opaque"). Added ErrorSetDecl, dispatched through the same _get_zig_container_name resolver ContainerDecl already uses. This alone dropped extra_classes from 10 to 1.
2. gitgalaxy/standards/language_standards.py: class_start widened with an optional \(/\* step-over between = and the packed/extern modifier, confirmed against two real corpus idioms: a parenthesized-wrapped anonymous struct (const lenAsc = (struct { ... }).lenAsc;) and a pointer-to-opaque handle type (const HMONITOR = *opaque {};, the standard idiom for an opaque OS handle).

The remaining extra_classes=1

Borrowed in bun/MimallocArena.zig is a confirmed tree-sitter-zig grammar limitation, not a GitGalaxy bug: the file uses Zig's newer #field private-field syntax, which the grammar can't parse (has_error=True on that node), so error recovery swallows the container declaration. Same shape as csharp's LanguageParser precedent in #1264.

--regenerate correctly refuses when it sees a raw extra_classes regression (0→1), even though found_classes improved from 0 to 630 in the same run. Per the doc's own documented escape hatch for this exact situation, I hand-blessed tests/tree_sitter_accuracy_baseline_zig.json to the measured numbers and explained why here, same as csharp's precedent.

missing_classes stays non-zero by design

found_classes 0→630 out of real_classes 642→656 (grew because the ErrorSetDecl fix exposed more real declarations). The missing_classes set (23→26 net) breaks down into four understood-but-not-chased shapes, written up as a new recurring-cause class in the checklist doc:

• Anonymous struct/enum nested inside a generic call argument (const Extra = List(struct { u32 });) -- not the direct declared value.
• Anonymous struct nested inside an array-element type (const targets = [_]struct {...}{...};).
• Anonymous type in type-annotation position, not initializer position (var op: enum {...} = .example;) -- recurred 6x in the corpus, the most promising follow-up candidate.
• A binary chain of named error sets terminating in a literal (const X = A || B || error{C};) -- would need meaningfully more regex complexity to skip safely.

None of these are GitGalaxy bugs or ground-truth bugs -- they're legitimate scope boundaries worth a dedicated follow-up decision rather than folding into this PR.

Result

found_classes 0→630, real_classes 642→656, extra_classes 0→1 (documented, not a bug). zig added to detector.py's _CLASS_START_NAMED_EXTRACTION_LANGS.

Golden master updated (both modes): bun/comptime_string_map.zig's class count 6→7 (the parenthesized-struct fix) and zls/mach/win32.zig's 158→159 (the pointer-to-opaque fix) -- both intentional, all other diffs are downstream numeric fallout of those two signal changes.

Test plan

• python tests/tools/class_start_diff.py --lang zig -- iterated until only the one confirmed tree-sitter parse-error case remained "extra"
• python tests/tools/tree_sitter_accuracy_audit.py --lang zig --ci / hand-blessed baseline after --regenerate correctly refused on the documented non-bug regression
• python tests/tools/tree_sitter_accuracy_audit.py --summary-table
• python tests/tools/update_golden_master.py --yes in both .crucible_venvs/{full_precision,zero_dependency} (confirmed each venv's editable install resolves to this checkout first)
• python tests/tools/crucible_check.py (both modes) -- clean, zero diff after regenerating
• python -m pytest tests/core_engine/ tests/extraction/ -- 6584 passed, 2 skipped, 11 xfailed, 3 xpassed
• python tests/tools/audit_check.py --regenerate -- ruff baseline regenerated (pure line-shifts from the new comment block only), mypy/dead-key/ast-accuracy all clean

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab