det: a float zero divisor is libcob's NaN; a phrase binds to the innermost open statement (#4675, #4676)

[squid-protocol]

Fixes #4675, Fixes #4676

Stacked on #4677 (-> #4654): merge those first.

Measured (GnuCOBOL 3.1.2 -std=ibm, 2026-10-08)

• COMP-1 / COMP-2 zero divisor (COMPUTE G = F / FZ, DIVIDE FZ INTO G, DIVIDE F BY FZ GIVING G or a decimal receiver): every receiver unchanged and the size error raised, with or without ON SIZE ERROR. Same as the decimal NaN (det runtime: a zero divisor without ON SIZE ERROR throws in the port; the oracle leaves the receiver unchanged #4655): the oracle converts floats to decimals. ON SIZE ERROR also fires for a decimal receiver too small for a float result.
• A zero divisor inside a larger float expression meets cob_decimal_align: (F/FZ) * 3 unchanged, but (F/FZ) * C, FM * (F/FZ) store 0 and 2 - (F/FZ) stores 2.
• Phrase scope: a NOT ON SIZE ERROR after a nested, unterminated COMPUTE belongs to the nested one; an END-COMPUTE closes the innermost COMPUTE; a stray phrase after a finished statement is a cobc syntax error.

Changes

• cobolrt/Hfp.divide: a zero divisor is the NaN (raises the size error) instead of an ArithmeticException; new Hfp.divideNested refuses by name a zero divisor inside a larger float expression (not replayed, C14).
• gen.py: float COMPUTE / ARITH take ON SIZE ERROR phrases (storeChecked on the value field), lifted receivers guarded against the NaN, phrase bodies generated in their own arithmetic mode (phrase_block; they inherited the statement's float mode before).
• det/stmt.py: _phrase_owner / _end_owner bind a phrase to the innermost unterminated statement that can take it and an END-verb to the innermost open statement of its verb. I did not touch stmt.parse's NOT rewriting (det parse: abbreviated 'OR NOT = x' relation; pin conditions holes 1-2 (#4681) #4683); my edit is the PHRASES / STATEMENT_ENDS branches of the node loop (the same function body, different branches: expect a textual merge at most).
• Docs: oracle_assumptions C14 extended (float rows, phrase scope).
• Shared functions touched: Cobol.nan / Cobol.raiseSize (private -> package, no behaviour change). Not touched: layout, Cobol.store / move (P-scaling det translator: a PIC SVPP99 COMP-3 item crashes with ValueError instead of a refusal by name #4669/det runtime: MOVE 1200 into a 99PP item throws in Cobol.store #4670).

Evidence

• tests/cobol_mainframe/test_det_size_error.py: float shapes H1..HB and nested shapes U1..U6 added to the end-to-end program (cobc vs port, both modes) plus a parse-level test. Before the change the end-to-end test fails (float phrases refused; nested shapes B2/B3/B7 differ), after it passes.
• Fuzz (/tmp/gitgalaxy-scratch/claude/det-4675/fuzz3.py: float items, zero divisors, nested size-error phrases 2 deep, terminated and not): 140 seeds, 104 comparable programs, 0 differences, the rest are tree-sitter "does not parse" holes on the nested shapes. The earlier det runtime: a zero divisor without ON SIZE ERROR throws in the port; the oracle leaves the receiver unchanged #4655 decimal fuzz (18 seeds) still equal.
• det_port.py check vs det runtime: a zero divisor leaves the receivers unchanged, as the oracle (#4655) #4677: all 68 ports "changed" only in the shipped runtime (Cobol.java modifiers, Hfp.java); no service differs. Re-proved mortgage-nbrvl and carddemo-acctview PROVED; mortgage-mpmt DIFFERS as on the base (C6 exponent hole).
• evidence_report.py --check clean. pr_gates.py --fast 5/5; --ratchets 7 pass, 0 fail, 0 skipped.

Not covered

• Float expressions with a zero divisor other than the whole statement value: run stops by name (follow-up filed).

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab