[squid-protocol]

Part of #4724 (readiness-scanner epic #4721).

What

• tests/tools/support_matrix.py: the support matrix for each burned estate. It reads committed files only, with no corpus and no scan:

  • the feature inventory fixture (Readiness scanner: feature inventory extractor — every COBOL/CICS/SQL feature, data type and option an estate uses, with program counts #4723);
  • the estate's evidence report (Evidence report: purchaser-facing levels and numbers per program and estate (no binary "proven") #4601): each program's level, its proving case, and its translate-only refusal or holes;
  • the rule catalogue cobolrt_rules.json (cobolrt: rule catalogue — an ID, IBM reference, option dependencies and register status for every runtime rule #4711, merged as cobolrt rule catalogue: an ID, IBM reference, options, register and calibration per runtime rule (Part of #4711) #5086 while this PR was in flight). Without it, every feature reads "not yet catalogued", as charter section 5 says;
  • the oracle-assumptions register.

  Each feature gets one of the issue's four statuses:

  • supported and proven: a program that uses it is at L2 or above; its cases and best level are shown;
  • supported, unproven: translated, and no refusal names it;
  • refused by name: with the refusal's owner;
  • unknown / never seen: only programs refused whole use it.

  Beside the status, it lists the conditions of "supported" from definition_of_done.md section 5 that the feature does not meet yet: modelled, catalogued, proven (L2), IBM-matched (an ASSUMED or DIFFERS rule means "uncalibrated"), calibrated on z/OS. For each estate it also lists:

  • the refusals with their owners;
  • system copybooks: a configuration gap ("IBM MQ copybooks not found: point the scanner at SCSQCOBC"; LE: SCEESAMP), not a translator limit (design point 2 in the issue comment);
  • runtime APIs not modelled: MQI, "translates, not provable", never "supported" (design point 3);
  • unknown features, for triage: the inventory's unknown buckets, unparsed programs, and keys that hold a sequence number.

  Usage: python tests/tools/support_matrix.py [ESTATE ...] [--json] [--catalogue PATH] [--out FILE].

• docs/language_status/refusal_map.json: 29 entries. Each refusal or hole text of the burned reports, and each missing copybook of the inventory, maps to one of:

  • an issue;
  • an agreed refusal family of the charter (charter keys, each checked to appear in definition_of_done.md);
  • an IBM product library;
  • an explicit "needs an issue" note.

  Force-added, like its sibling JSON files, because of the repo-wide *.json ignore.

• tests/cobol_mainframe/test_support_matrix.py (19 tests):

  • Synthetic estate: each status; each refusal kind (an issue, a charter family, a system copybook for both MQ and LE, an estate copybook, unmapped); the MQ API section; the catalogue absent, and present with an ASSUMED rule, a compile-option join and a word boundary (NUMVAL is not NUMVAL-C); the triage lists; deterministic render.
  • Committed data, the acceptance ratchet: every refusal and hole of the six burned reports matches exactly one map entry. The map entries are well formed. Every burned estate renders. The JSON CLI works. Every rule the committed catalogue joins exists.

• Docs: a "Support matrix" section in feature_inventory.md, and a pointer in charter section 5.

Matrix today (rule catalogue on main)

estate	proven	unproven	refused	unknown	no catalogue rule	refusal kinds
CardDemo	162	12	1	70	186/245	10
cics-async credit card	46	5	1	0	31/52	1
CBSA	183	29	2	0	155/214	3
GenApp	161	0	0	0	116/161	0
DBB mortgage	55	36	10	0	68/101	10
zECS	33	202	3	0	198/238	3

• CardDemo's 70 unknown features come from its 3 MQ programs. The inventory parses those only by fallback, and that fallback reads sequence numbers as words (for example RETRIEVE 01327507). The matrix lists them for triage.
• "No catalogue rule" is high because the catalogue covers the runtime's data-movement rules. Most CICS command names, for example, have no rule whose title names them. The join is family plus title words: a stated heuristic until the catalogue carries feature keys.
• The levels and refusals come from each report's survey (0d74a18). A gap fixed since then shows until the nightly evidence refresh. For example, CBSA DELCUS READ/DELETE TOKEN (det/CICS: READ / REWRITE / DELETE / UNLOCK TOKEN (update tokens; makes CBSA whole) #5078, fixed by CICS: file update TOKEN (READ UPDATE / REWRITE / DELETE / UNLOCK TOKEN); DELCUS translates whole (Part of #5078) #5090) and BANKDATA CEEGMT/CEEDATM (det: LE date/time services (CEEGMT, CEEDATM, CEESECS, CEEDATE, CEEISEC, CEEQCEN) so CBSA BANKDATA translates whole #4949, modelled by det: LE date/time services CEEGMT, CEEDATM, CEESECS, CEEDATE, CEEISEC, CEEQCEN (Part of #4949) #5002).

Class

Readiness-scanner support status per inventory feature, joined from committed tool outputs. This PR covers every member: the four statuses; the refusal kinds (issue, charter family, system copybook, unmodelled API, needs an issue); the catalogue absent, present and uncalibrated; the compile-option join.

Refusals owned by #5106 (the consolidated siblings issue): ACCEPT FROM SYSIN; CALL MVSWAIT; a COMMAREA DTO property the port cannot convert (List<String>); files with no repository in the generated project; LE condition handling (CEEHDLR / CEEHDLU, also as the unmodelled API); CALLs to sibling programs of a multi-program source; procedure pointers (SET TO ENTRY); SOAPFAULT CREATE. Acceptance is now enforced: support_matrix.owned() requires each entry to have an issue, an agreed refusal family or a system library, and an unmodelled API to have its issue. The matrix flags any other entry as an acceptance gap, and test_every_refusal_map_entry_has_an_owner / test_no_burned_estate_has_an_unowned_refusal fail on it.

Sibling gaps (not this PR):

• Publishing the page to generated. It needs a generated-docs.yml edit, which was out of bounds here: the publish step lists its pages, and test_generated_branch.py pins that list to PAGES. Until then, render it with --out.
• SYSLIB-style system library paths for COPY resolution (design point 1): a scanner input, customer side (Readiness scanner: deployable offline package — runs on the customer's side, no source leaves #4726).
• Stand-in copybook use (CEEIGZCT, det: a stand-in for LE's CEEIGZCT (feedback codes for CEEDAYS) so CBSA CRECUST translates whole #4884, design point 4): the inventory does not record it, so the matrix cannot show it.
• No CALL-target category in the inventory. An unmodelled API is detected only through its copybook (CMQ*) today.
• Inventory CICS keys hold sequence-area numbers in fallback-parsed programs (feature_inventory.fallback_parse does not strip them).
• No feature keys in cobolrt_rules.json. Until it has them, the join is a title heuristic.

Older CICS estates to consider (not cloned)

IBM's own CICS TS product samples, from the customer's SDFHSAMP: the FILEA sample programs (DFH$AMNU / DFH$ALLB family), the NACT account sample from the CICS Application Programming Primer, and the CICS catalog manager example (DFH0XCMN ...). All three are IBM-licensed and would be scanned customer-side (#4726), never shipped. I did not open the blind estate-4 candidate list.

Gates

• pr_gates.py --fast (heavy-run, JDK 17 env): 6/6 pass (lint with CI's ruff 0.16.10, audits, secrets, xray, gen-docs, gauntlet).
• pr_gates.py --ratchets: 7/7 pass, 0 skipped (cics-spec, ports, port-surface, estate, fact-crosscheck, completeness, ground-truth). No engine, translator or parser change, so no golden bless.
• audit_check.py: all clear. Support-matrix tests: 19 pass (head a90312d, pr_gates --fast 6/6 re-run), run with plain pytest from /tmp.

Cross-repo

None.

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab

[squid-protocol]

No description provided.

[squid-protocol]

No description provided.