fix(refraction): forge parser fixes (#3203 1-4, #3204, #3205, #3206, #3212 determinism)

[squid-protocol]

Fixes the refraction tools' own parsers found by the #3120 differential. Tool-side only: no engine file changes, so no golden-master bless and no rosetta impact.

Closes #3204
Closes #3205
Closes #3206
Part of #3203 (defects 1–4; defect 5, the SECTION model, stays open)
Part of #3212 (the determinism half; the snapshot test stays open)

What changed

issue	fix
#3204	extract_lineage walks the OPEN operand list and switches mode at each INPUT/OUTPUT/I-O/EXTEND keyword.
#3205	analyze_cobol_intent counts EXEC CICS / EXEC SQL by their opening keywords. The unbounded lazy EXEC….*?END-EXEC\. over the joined file is gone.
#3206	cobol-refractor never writes the target. A patched program goes to <clean room>/00_patched_source/<relpath>, and the forge tools read that copy. patch_lexical_content() is the pure core; patch_lexical_traps() takes an optional dest.
#3212	IR dumps serialise sets sorted (_ir_to_json).
#3203 1–4	Paragraph headers must start in Area A (cols 8–11). END-*/GOBACK/EXIT/CONTINUE are never headers. Sequence fields in cols 1–6 are tolerated. Copybooks resolve across the whole repository, nearest first, and never to a member with a PROGRAM-ID.

One extra defect of the same shape turned up while scoring. After the split on PROCEDURE DIVISION, the rest of the header line ( USING PARM-BUFFER.) read as the entry paragraph on 18 CBSA programs. That is how BANKDATA's A010 came out dead (the #3203 comment). paragraph_headers now skips that line.

refraction_differential.py and cobol_answer_key.py now call the graveyard's own paragraph_headers / find_copybook instead of copies of its old regexes, so they measure the forge as shipped.

Scored against the answer key (#3210), forge before → after

field	zopeneditor-sample	CBSA
units	P 58/75 → 58/58	P 452/729 → 452/452
dead (non-trivial)	17 false → 0 claimed	P 5/571 → 5/328
copybook paths	R 0/14 → 12/14	P 0/29 · R 0/114 → P 99/99 · R 99/114
inputs	P 6/12 → 6/6	—
outputs	R 0/6 → 6/6	R 0/1 → 1/1 (BANKDATA)

What remains on CBSA is all defect 5. The 228 missing units are all SECTIONs, and all 323 remaining false dead claims are in programs that use sections (fall-through isn't modelled). EXEC counts now match a line-level count on all 31 CBSA programs (BNKMENU 62, XFRFUN 77, BNK1CAC 37).

Recorded in a new update section of docs/refraction_engine_differential.md.

Verification

• tests/cobol_mainframe: 81 passed. 12 of the 13 new tests fail against origin/main's code; the 13th is an existing test matched by the -k filter.
• audit_check.py: all clear (ruff, mypy, dead-key, ast-accuracy).
• End to end: two cobol-refractor runs directly on a CBSA clone left git status clean and gave identical IR dumps and JCL. On a zopeneditor copy with an injected NEXT SENTENCE, the target stayed clean and the patched copy landed in 00_patched_source/COBOL/SAM2.cbl.

Found, not fixed here

Clean-room outputs are keyed by file stem, so zopeneditor's multiroot/sam/SAM2.cbl overwrites COBOL/SAM2.cbl's IR dump, JCL and schema. This predates this PR (00_patched_source mirrors relative paths, so it isn't affected).

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab