fix(cobol): func_start no longer mistakes LE service calls for paragraphs

[squid-protocol]

Closes #2538.

Root cause

CEE3DMP./CEEMOUT. (Language Environment runtime diagnostic calls) sitting alone
on a line indented into COBOL's Area B were counted as paragraph headers by
func_start — the reserved-word negative-lookahead shield had no entry for them,
and the post-sequence-area whitespace slot ([ \t]*) tolerates any horizontal
indent to support genuinely free-format source, so it can't distinguish Area A
(paragraph names, columns 8-11) from Area B on its own.

Was this deliberate?

No. #2489 (cited in the issue) is the previous func_start sequence-area
hardening, but it only narrowed the vertical-whitespace slot ([ \t\n]* →
[ \t]*) to fix a debug-line name-mangling bug — unrelated to column position.
Confirmed via git log that Area A/B has never been discussed or implemented in
this file's history before this issue.

Fix

Added CEE3DMP|CEEMOUT|CEEDUMP to func_start's existing reserved-word shield —
the same low-risk, precedent-backed fix pattern already used repeatedly in this
exact file (CONTINUE for #1890's sibling class, LOCAL-STORAGE,
SOURCE-COMPUTER/OBJECT-COMPUTER). These three tokens are already a recognized
class elsewhere in cobol.py (the telemetry rule), so this reuses an existing
categorization rather than inventing a new one.

Considered and deferred: a general Area-A column anchor (the issue's other
option). The existing optional 6-char sequence-area consumer is itself a loose
heuristic — [0-9a-zA-Z \t]{6} happily matches 6 leading spaces in a free-format
file too, so "did the sequence-area group engage" doesn't reliably distinguish a
genuine fixed-format line from a free-format one. A correct structural fix needs
file-level fixed/free-format detection, which is a materially bigger change than
this specific bug warrants — left as a follow-up if the general class (any bare
token in Area B, not just these three) needs closing later.

Verification

• Reproduced the exact FP locally: 5 func_start matches for 3 real paragraphs
  (CEE3DMP./CEEMOUT. both miscounted); confirmed the fix brings it back to 3.
• Added cases to tests/extraction/languages/test_cobol.py (invalid-match cases)
  and test_cobol_strict.py (reserved-verb exclusion sweep).
• test_cobol.py + test_cobol_strict.py: 148 passed.
• tests/core_engine/test_language_standards_strict.py: 21 passed.
• tests/tools/audit_check.py: clean (ruff/mypy/dead-key/ast-accuracy).
• tests/tools/crucible_check.py (both full_precision and zero_dependency
  modes, full corpus): PASS, zero diffs — these tokens aren't present in the
  existing golden-master/crucible corpus, so no other output changed.

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab