COBOL engine: a group's USAGE applies to its elementary items (#4525)

[squid-protocol]

Closes #4525

Problem

A USAGE on a group (01 COMP-VARIABLES COMP.) was not applied to the elementary items under it. The engine sized them as DISPLAY, so their sizes and every offset after them were wrong. CardDemo CBSTM03A: COMP-VARIABLES came out at 16 bytes (should be 8) and WS-TOTAL-AMT at 11 bytes (should be 6).

Fix (IBM Enterprise COBOL rule: an item with no USAGE takes the USAGE of the nearest enclosing group)

• Extractor (mainframe_boundary): record_data.usage now holds the inherited usage. This extends the Mainframe byte level: raw EBCDIC code pages (037/1047/277/278/930/939, FB80 splitting) and DBCS COBOL (PIC N/G, SO/SI) #3816 NATIONAL / DISPLAY-1 rule to every usage. A nested group's own USAGE overrides an outer one. An inherited DISPLAY is not recorded, but it still blocks an outer usage. 66/88 entries keep their Mainframe byte level: raw EBCDIC code pages (037/1047/277/278/930/939, FB80 splitting) and DBCS COBOL (PIC N/G, SO/SI) #3816 shape (no storage). A 77 belongs to no group.
• GalaxyIR (record_layout, _storage_spans): the enclosing usage is passed down the walk, so COPY-expanded members also pick up the usage of the group they are COPYed under. A picless usage (POINTER/INDEX/COMP-1/2) on a group no longer makes that group elementary.
• Schema forge reader (data_entries) now inherits too. Without this, the refraction differential flagged CBSTM03A's WS_TOTAL_AMT.
• PL/I is not affected. It uses a separate path (_pli_record_layout / pli_mapping), and PL/I data attributes are not inherited from a structure.
• REDEFINES and OCCURS members inherit like any other member.

Ratchets

• Fact cross-check ledger (fact_crosscheck.py update, unscoped): all 14 engine-group-usage-not-inherited disagreements are gone, the cause is dropped, and 0 are untriaged. The translator adapter's usage channel now reads the translator's effective usage, which is what the engine now records. The old as-written snapshot wrapper is removed.
• Golden masters: blessed with crucible_check.py --update --yes on both legs. The drift is 2 NIST CCVS85 record layouts (NC1354.2, NC2454.2, COMPUTATIONAL groups). scope_check.py --expect cobol passes on both modes. No timeouts.
• Ground truth, completeness, estate baseline and ports: PASS, unchanged.
• Committed ports/evidence: a sweep of all mainframe corpora, estate-crucible and cics-crucible finds group-level non-DISPLAY usages only in CBSTM03A, which has no port or evidence record. Nothing needs re-proving.

Tests

tests/cobol_mainframe/test_group_usage_4525.py covers:

• the CBSTM03A lines verbatim, plus the real corpus file
• COMP, COMP-3 and POINTER groups
• a nested COMP-3 group and a DISPLAY override
• an item's own USAGE
• OCCURS and REDEFINES
• 88 and 77 entries
• a COPY member under a COMP-3 group
• agreement with the storage spans

test_dbcs_items had pinned the old "not inherited" behaviour and now asserts the inheritance.

🤖 Generated with Claude Code

https://claude.ai/code/session_01W3KVtYHQp8KsfKcHsdiPbR

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab