feat(core-engine): yacc's %union is a real class_start, with the allowlist entry that makes it safe (#2644)

[squid-protocol]

Closes #2644. Part of #2669.

Implements the fix shape from the issue, plus the blocker its own follow-up comment identified —
both halves in one PR, because either alone is a regression.

The gap

yacc.py wired class_start: None, documented as "a grammar file has no object/type concept".
True for OOP-style classes, but Bison's %union declares a real compound type: the C union
spanning every rule's semantic value ($$/$1, which args already counts). Same mapping the
engine already makes for Fortran's TYPE … END TYPE, COBOL's PROGRAM-ID, assembly's struc.
And it is core grammar syntax, not incidentally-embedded C — internal_discriminator already
lists union among the %-directives used to identify a file as yacc.

"class_start": re.compile(r"^[ \t]*%union\b(?:[ \t]+([a-zA-Z_]\w*))?", re.M),

Optional group 1 for bison's rarer named-tag form (%union name {); the common anonymous form
resolves to Anonymous_Class, the path assembly's own no-name class_start takes. No group 2 —
a union has no inheritance parent for detector.py's group-2-is-parent convention to misread.

Why the allowlist entry is the load-bearing half

detector.py's named-class extractor consults a language's own class_start only for
languages in _CLASS_START_NAMED_EXTRACTION_LANGS. yacc was absent, so wiring the rule alone
drops it onto the legacy generic fallback (class|struct|interface|trait|enum), which reads every
struct foo declaration in a grammar's embedded C actions as a class:

file	%union	honest answer	fallback (rule wired, no allowlist)	this PR
yacc/freebsd/config.y	1 (line 1)	1	17	1
yacc/freebsd/jailparse.y	1 (line 45)	1	9	1
cobol/gnucobol_internals/parser.y	0	0	56	0
cobol/gnucobol_internals/scanner.l	0	0	27	0

109 phantom classes across the four real grammar files, against a true total of 2. The rule
without the allowlist entry would have been worse than the None it replaced, which is why
#2644 was marked blocked on a slicer-layer companion change — this is that change.

Verification

yacc is tree-sitter-blind, so tree_sitter_accuracy_audit.py cannot check it (it is not in
NODE_MAPS), and ctags cannot corroborate classes either: ctags --list-kinds-full=YACC
exposes exactly one kind, l (label), which is why CTAGS_CLASS_KINDS["yacc"] is empty on
purpose and the §9 tri-comparison covers functions only. Verified instead by direct source
cross-check against all four real grammar files — the same standing abap, cobol, jcl and sqlite
have on this allowlist. Result: 100% precision, no false positives, and an honest zero where a
grammar declares no union.

• Golden masters re-blessed, both modes. The diff is exactly two lines in each:
  Class/Entity Declarations 0 → 1 on config.y and jailparse.y. Nothing else moves — no
  structural mass, no risk exposure, not even a topological coordinate.
• pytest tests/ — 7583 passed. The yacc gauntlet's CLASS_CASES were empty (the rule was
  None) and are now populated, with the invalid cases that matter here: a bare C union/struct
  in action code, mid-line and commented-out occurrences, and %unionize against the \b guard.
  Two detector-level tests pin the two halves together (one %union amid C structs → exactly one
  class; a grammar with no %union → none), plus ReDoS coverage for the new quantifier.
• ruff_audit.py --ci, mypy_audit.py --ci, dead_key_audit.py --ci — clean against baseline.
• tri-comparison-audit does not cover yacc (baselines exist for typescript, javascript and zig
  only), and tree-sitter-accuracy-audit cannot see it at all.
• docs/language_status/yacc.md §1/§3/§4/§6/§7/§8 updated — including the stale key count, which
  read 31/47 and is now 33/48 measured.

Corpus follow-up (keyword-rosetta)

The yacc/class_start n/a cell becomes plantable: the rosetta corpus can now add a %union probe
to data/yacc/*.y and update expected_signals.json. Screened against yacc's full rules dict, the
verbatim config.y:1-6 shape fires nothing but class_start (no (, ,, =, &, <, void,
const, static, comment markers, or line-ending :). Not included here — that is a corpus PR
against merged engine main, per the no-pins flow.

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab