Fix #3992: no-JCL batch controller names an undefined <File>DTO

[squid-protocol]

Fixes #3992

Root cause

JCL isn't actually involved. base_intent.files_requested comes from analyze_cobol_intent's SELECT/ASSIGN anchor (cobol_jcl_forge._FILE_ASSIGN_ANCHOR), not from JCL. That anchor only read a word operand. So:

• ASSIGN "TEST-FILE" / ASSIGN "./TEST-FILE" (the form used by GnuCOBOL, opensourcecobol4j and Micro Focus) matched nothing. files_requested came back empty, the program was classified transactional, and the controller bound @RequestBody TestFileDTO. No generator ever writes that class, so compilation failed with cannot find symbol.
• ASSIGN TO "./TEST-FILE" backtracked past the optional TO and read the keyword TO itself as the DD name. That compiled, but produced a wrong @RequestParam("toFile") and a //TO DD in the zero-trust JCL.

The engine's own reader (mainframe_boundary._SELECT_ASSIGN) already accepts literals and SELECT OPTIONAL, so lineage.inputs had TEST-FILE while files_requested was empty.

Approach: option (a) plus a general guard

1. cobol_jcl_forge: the SELECT anchor now reads a quoted literal (bounded to 1024 chars), SELECT OPTIONAL, and never takes TO as the operand. This matches the engine's reader. A non-CICS program doing file I/O is batch, so it gets the MultipartFile upload signature, the same way CardDemo batch programs do.
2. api_contract_forge, as a guard for any SELECT form the intent reader still misses:
   • A non-CICS program with lineage inputs is batch even when files_requested is empty. Its input files become the upload params.
   • The transactional branch no longer names a <File>DTO. No generator writes a per-file DTO (a CICS program's COMMAREA contract comes from its skeleton via forges.cics.controller). The inputs are listed in a // Input files: ... comment for the porter. Two @RequestBody on one method was never valid Spring anyway.
   • A DD name from a literal ("01.DAT", KUNDE§NR) is made a legal Java identifier (java_legal_chars + file prefix if it can't start one). Legal names are unchanged.

Why not only option (b)? Keeping the program transactional and dropping the parameter would compile, but it would describe a file-processing batch program as a request/response endpoint and leave the JCL reading TO. Option (a) is the semantically right classification. The guard means the controller can't name an ungenerated class again, even if the intent reader misses a SELECT.

Output impact

• CI corpora (CardDemo, CBSA, GenApp, zopeneditor, zECS, DSF): generated Java is byte-identical to main (java_target_matrix.py --same-as, default config, plain and --scan paths). None of them uses literal ASSIGN or SELECT OPTIONAL.
• test_golden_forge.py's golden controller changed. Its mock IR was exactly the Generated REST controller references an undefined <File>DTO for a batch program without JCL: jp-compat still does not compile after #3957 #3992 shape (non-CICS, inputs, no SELECTs) and it codified the two undefined @RequestBody ...DTO params. It is now the batch upload signature.

Estate: opensourcecobol4j jp-compat (mvn -o compile, JDK 17)

	before (origin/main)	after
compile	FAILED: 10 controllers cannot find symbol: class TestFileDTO	OK
controllers	36	36
@RequestBody TestFileDTO	10	0
@RequestParam("toFile") (TO misread)	15	0
upload params	toFile×15, sys01DaDkS*×4	testFileFile×28, inputTxtFile×2, nowhereFile×2, sys01DaDkS*×4

The zero-trust JCL for the literal-ASSIGN programs now declares //TESTFILE DD with the right DISP instead of //TO DD ... NO EXPLICIT OPEN INTENT. The analyze.py findings are unchanged.

Tests

• New tests/cobol_mainframe/test_no_jcl_batch_controller.py:
  • The intent reader handles literal, TO literal, OPTIONAL and word forms, and never reads the TO keyword as a name.
  • Controller unit tests: no-SELECT non-CICS is batch, a CICS program names no DTO, literal names become legal identifiers.
  • End to end on a no-JCL estate: scan, refract, generate. It asserts every controller names only classes the run generated, and that mvn -o compile succeeds (skipped without Maven + JDK 17 or an offline-resolvable repo).
  • 13 of the 17 fail on main.
• tests/cobol_mainframe + tests/tools_recorders: 1338 passed, 32 skipped.
• ruff_audit --ci, mypy_audit --ci, dead_key_audit --ci: no new findings. tool_regex_redos.py --ci: 0 offenders. ruff format is clean on the touched files.

Not addressed

cobol_dag_architect._SELECT (the forge lineage used without --scan) has the same ASSIGN TO "x" → TO backtrack. Its code_view blanks literal contents, so it can't read the literal name. The controller guard keeps that path compiling.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EC8FWfn3aPmsUPnRgKNVup

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab