[squid-protocol]

Part of #7524, #7709

Checklist: done [c:9bd1fd57]

What

The JSON GENERATE / PARSE translation is merged (X76 / X77, fixture tests, oracle-unbacked; #7851 ... #7946). This is the issue's last acceptance line, "re-survey: the 2 census rows drop", on the committed evidence. At main 945e52b79: cics_census.py survey --baseline, then gap_prevalence.py --survey <before-b link> --census-survey <before-nb link> --census-corpora <census clones> --write, then cics_census.py history append. Counts only; no census source read, copied, edited or committed.

Result (tool only)

• Whole scope: 114 / 127 translate whole, non-burned 9 / 21, 8 refused whole. 1 copy member listed, not counted.
• Census-only, basis 39: 12 of 39 whole, 36 of 39 accounted. Coverage if closed: top 1 -> 13, top 3 -> 24, top 5 -> 32, top 10 -> 37 (upper bounds).
• JSON GENERATE: not modelled and JSON PARSE: not modelled are gone from the committed census summary and page (they read 1 estate / 1 program each before). No JSON row remains in the survey.
• New row exposed (not JSON): EXEC CICS not a data area: '…', 1 census program, only gap (it sat behind WEB CONVERSE POST, closed by det: WEB CONVERSE takes the bare method keywords (Part of #7525, #7709) #7942). It scrubbed to EXEC: other / unowned, which fails test_committed_summary_has_no_unowned_row_beyond_the_named_exemptions_7847. Fix here: a name-free construct key EXEC CICS: operand is not a data area (owner COBOL translator: CICS command breadth (START/RETRIEVE, channels/containers, GETMAIN, LOAD, CONVERSE, INQUIRE, ASSIGN, counters, HANDLE ABEND PROGRAM) #4270 via the existing ^EXEC CICS pattern), with scrub tests for the literal and the Length of <name> spellings. The operand text is source, so only the class is kept.

Class / Sibling gaps

Class: stale committed census evidence after merged translator slices, plus a census gap key that collapses to the unowned EXEC: other catch-all. Members handled: JSON GENERATE row, JSON PARSE row, WEB CONVERSE POST row (all dropped by the regeneration), and the not a data area key (scrubbed and owned). Test covers both literal and length-expression spellings.
Sibling gaps: (1) EXEC CICS operand where the det port wants a data area (cics.py ref): 1 census program, no burned program; which option it is cannot be read without census source, so it needs a fixture (a cics-crucible program or fixture with a literal and a LENGTH OF operand), scope sweep/frontier. (2) JSON shapes IBM's text leaves silent (elementary source or receiver, CONVERTING, COMP-1/2 floats #7947 open, PARSE into tables, JSON-STATUS after an exception): no census row, no burned program uses JSON. (3) XML PARSE / GENERATE: holes by name (#7759), no census row.

Gates

pytest tests/tools/test_gap_prevalence.py test_cics_census.py test_blocker_report.py pass; pr_gates --fast 9/9; audit_check.py all clear. No engine change, so no golden or ratchet run.

Lessons for agents

• gap_prevalence.py --write with no --survey keeps the committed burned summary; pass a dir holding only a before-b symlink to refresh the burned half and one holding only before-nb for the census half.

🤖 Generated with Claude Code

[squid-shepherd]

No description provided.

[squid-shepherd]

No description provided.