rosetta[cobol] #2567 sweep: 6R/4A → 3R/4A — #2537 re-baseline, shell re-authoring, args morphology ledger

[squid-protocol]

Summary

The cobol leg of epic squid-protocol/gitgalaxy#2560, worked per the rosetta-language-sweep
skill against tracking issue squid-protocol/gitgalaxy#2567. Live table (which supersedes the
issue title's 4R/8A — the post-#2618 api re-baseline had already moved keyword_hits and
risk_api_exposure): 6 red / 4 amber → 3 red / 4 amber, and every residual is now ledgered
or pinned to a named cross-cutting issue — the epic's close criterion.

The five-bucket classification

1. Real engine bug — fragile_debt +200% / planned_debt +100% = gitgalaxy#2537
   (debt keywords matching inside hyphenated identifiers: HACK-LEVEL, PROBE-TODO).
   Fixed upstream by gitgalaxy PR #2622; cobol c.cpy re-baselined 3→1 / 2→1 and scheme
   c.scm 2→1 here (scheme reproduced the leak via its probe-todo symbol).
2. Missing rule w/ morphology — none this sweep.
3. Corpus authoring gap — comment_lines 14 vs median 30 (−53%, lowest of all 46):
   cobol was the only corpus language authored with zero blank lines (fortran/ada
   siblings carry ~12). Shells re-authored with idiomatic spacing + neutral banners
   (debt-keyword-safe; signal placement unchanged) → 33, in-band.
4. Intended morphology, ledgered — args 1 vs 13: NEW validated entry
   cobol-args-clause-match-morphology. Bucket-3 test run first (jcl PARM= precedent):
   verified empirically that the file-level metric counts USING/RETURNING clause matches,
   not names (a 13-name PROCEDURE DIVISION USING still records struct_args 1), and every
   per-probe alternative (CALL/ENTRY/INVOKE … USING) is also an api keyword — overlap-cost
   rejection. Also globals +50% (WORKING-STORAGE SECTION, already ledgered; keeping the
   rule was confirmed as deliberate design).
5. Median inflation / cross-cutting — state_mutation 2 vs 6: cobol is the honest one
   (planted=2), median inflated by #2546 ×3 flux. high_risk_execution +50% / io +33%:
   the a.cpy string decoy, blocked on #2535's shielding design question. §3 risks are
   downstream shadows (risk_cognitive_load −57%, risk_safety_score +42% remain until
   #2535/#2545/#2546 land).

Bias artifacts regen note

docs/bias_data.json was last committed before engine #2618 (orphan/duplicate census fix)
merged — PR #8 re-baselined the 13 api manifest cells but not the cache. This regen folds
that drift in (sqlite/shell/livecode/lua/matlab/ruby risk_api_exposure / risk_tech_debt /
keyword_hits move). Verified the #2537 guard can only touch fragile_debt/planned_debt: the
two GLOBAL patterns feed exactly those keys in all 45 consuming languages.

Gates

verify_language.py cobol PASS (75 assertions), scheme PASS (72), both against the engine
PR's build (ENGINE_REF=pull/2622/head).

Cross-repo

• Companion engine PR: fix(standards): guard debt patterns against hyphenated identifiers (closes #2537) gitgalaxy#2622 (closes #2537). Its rosetta-audit
  check fails until this PR merges — expected, that's the pinned gate.
• ENGINE_REF is set to pull/2622/head so this PR's gates run against the engine fix now.
• Merge order: #2622 merges first (after this PR is green against it) → this PR restores
  ENGINE_REF=main and merges → gitgalaxy bumps KEYWORD_ROSETTA_REF to this PR's merge SHA.
• Tracking: rosetta[cobol]: cross-language consistency — 4 red / 8 amber metrics, 5 ledgered shapes gitgalaxy#2567 gets the sweep summary comment; the
  docs/language_status/cobol.md §10 capstone lands as a separate gitgalaxy docs PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BsVATdyMPhUoAUNBKbMUVi