rosetta[cobol]: cross-language consistency — 4 red / 8 amber metrics, 5 ledgered shapes

[squid-protocol]

Part of #2560 — per-language tracking for the keyword-rosetta control corpus (built for #1096). Planted intent is identical in every language, so every deviation below is measured bias vs. the 46-language median, validated in the corpus's deviation ledger. 🔴 = beyond ±50% of the cross-language median, 🟡 = ±25–50%.

Triage order (per the epic): structure detection first, then signal/keyword rules, then risk-score consequences — §3's numbers are downstream of §1–§2 and re-baseline as those fix. Cross-cutting engine issues (e.g. #2535/#2546) fix many languages at once; this issue tracks this language's residual after each cross-cutting fix lands, and closes when the language sits inside the ±25% band on every comparable metric (or every remaining deviation is ledgered as intended morphology).

1. Structure detection (fix first)

• 🔴 args: 1 vs median 13 (-92%)

2. Signal / keyword deviations

• 🔴 fragile_debt: 3 vs median 1 (+200%)
• 🔴 planned_debt: 2 vs median 1 (+100%)
• 🔴 state_mutation: 2 vs median 6 (-67%)
• 🟡 comment_lines: 15 vs median 30 (-50%)
• 🟡 globals: 3 vs median 2 (+50%)
• 🟡 high_risk_execution: 3 vs median 2 (+50%)
• 🟡 io: 4 vs median 3 (+33%)
• 🟡 keyword_hits: 178 vs median 247.5 (-28%)

3. Risk-exposure consequences (downstream)

• 🟡 risk_api_exposure: 6.81696 vs median 5.31104 (+28%)
• 🟡 risk_cognitive_load: 8.11636 vs median 15.1444 (-46%)
• 🟡 risk_safety_score: 53.1215 vs median 37.297 (+42%)

Ledgered shapes touching cobol

defect	type	issue	evidence	summary
census-requires-git-tracked	engine-semantic	—	—	GalaxyScope's census enumerates git-tracked files only; an untracked folder scans as '0 files mapped' with no per-file warning
cobol-working-storage-globals	keyword-overlap	—	(notes)	cobol globals pattern includes WORKING-STORAGE\s+SECTION as an alternative — every fixed-format program with a working-storage section records globals >= 1
cobol-hyphen-identifier-debt-leak	upstream-bug	#2537	c.cpy	Debt rules are whole-text, not comment-anchored, and '-' is a regex word boundary: HACK-LEVEL (data item, 2 code occurrences) counts fragile_debt, PROBE-TODO (paragraph …
cobol-area-b-bare-token-paragraph	upstream-question	#2538	(notes)	A bare token + period in Area B (CEE3DMP
string-literal-selective-shielding	upstream-question	#2535	a.cpy	Keywords inside one string literal ('if eval fails, try open') count toward branch (+2), safety (+1), io (+1) but eval does NOT count toward high_risk_execution (routes …

Decoy outcomes (file-level evidence)

• main.cbl · never ALTER this flow and no PERFORM loop (comment): stripped: contributed 0
• a.cpy · DISPLAY 'IF ALTER FAILS TRY OPEN AGAIN' (string): ALL counted: IF (branch +1), ALTER (high_risk +1), OPEN (io +1) — cobol does NOT shield the high-risk family in literals, unlike python/javascript. Ledger: string-literal-selective-shielding; gitgalaxy#2535 comment
• c.cpy · 77 HACK-LEVEL PIC 9. / DISPLAY HACK-LEVEL. (code): COUNTED (+2): hyphen is a word boundary so \bHACK\b matches inside the identifier, both occurrences; the PROBE-TODO paragraph name likewise adds planned_debt +1. Ledger: cobol-hyphen-identifier-debt-leak; gitgalaxy#2537

Reproduce / evidence

• Corpus folder: data/cobol/ — expected_signals.json carries the shape-by-shape accounting
• Gate: GALAXYSCOPE_BIN=<venv>/bin/galaxyscope python tools/verify_language.py cobol in the corpus repo
• Full context: findings_by_language.md#cobol · bias_report.md · bias chart

[squid-protocol]

Sweep complete (rosetta-language-sweep): 6🔴 / 4🟡 → 3🔴 / 4🟡

Worked per the corpus's rosetta-language-sweep skill. Note the live language_deviations.py
table had already superseded this issue's title (4🔴/8🟡): the post-#2618 api re-baseline moved
keyword_hits and risk_api_exposure in-band and re-banded comment_lines/risk_cognitive_load
to red before this sweep started — the sweep's true starting point was 6🔴/4🟡.

The five-bucket classification

1. Real engine bug — fixed. fragile_debt +200% / planned_debt +100% = #2537: the shared
debt patterns' bare \b(...)\b + --as-word-boundary matched debt keywords inside hyphenated
identifiers (HACK-LEVEL ×2, the PROBE-TODO paragraph name; scheme's probe-todo symbol
reproduced it). Fixed by PR #2622 (hyphen-adjacency guards on GLOBAL_FRAGILE_DEBT /
GLOBAL_PLANNED_DEBT; golden masters re-blessed — the same FP class died corpus-wide, e.g. an
SSN mask xxx-xx-xxxx counted as XXX and a roles-path-bug directory name counted ×4).
cobol → 1/1 exact, scheme planned_debt → 1. Re-baselined in keyword-rosetta PR #10.

2. Missing rule w/ morphology. None — cobol's rule surface was already the registry's deepest.

3. Corpus authoring gap. comment_lines 14 vs 30 (−53%, lowest of all 46): the cobol shells
were the only corpus language authored with zero blank lines (fortran/ada siblings carry ~12;
blanks are legal COBOL). Re-authored with idiomatic spacing + neutral banners (signal placement
unchanged) → 33, in-band (PR #10).

4. Intended morphology — ledgered.

• args 1 vs 13: new validated entry cobol-args-clause-match-morphology. Bucket-3 test run
  first (the jcl PARM= precedent): verified empirically the file-level metric counts
  USING/RETURNING clause matches, not captured 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 (jcl DISP= precedent).
• globals +50%: WORKING-STORAGE SECTION in the globals rule — deliberate shared-state
  design, already ledgered (cobol-working-storage-globals); keeping the rule confirmed.

5. Median inflation / cross-cutting — cobol is the honest one.

• state_mutation 2 vs 6: matches planted intent exactly; median inflated by state-flux ×3 branch-context weighting: undocumented, and phantom-triggered by branch keywords inside string literals #2546's ×3 flux.
• high_risk_execution +50% / io +33%: the a.cpy string decoy (ALTER/OPEN in a DISPLAY
  literal) — blocked on prism/detector: string-literal shielding is selective by signal family (if/try/open count inside a literal, eval does not) #2535's selective-shielding design question.
• §3: risk_cognitive_load −57% / risk_safety_score +42% are downstream shadows; re-baseline
  as prism/detector: string-literal shielding is selective by signal family (if/try/open count inside a literal, eval does not) #2535/kotlin: branch rule counts 'return' — inconsistent with java/swift/c siblings #2545/state-flux ×3 branch-context weighting: undocumented, and phantom-triggered by branch keywords inside string literals #2546 land.

Artifacts

• Engine: PR fix(standards): guard debt patterns against hyphenated identifiers (closes #2537) #2622 (closes fragile_debt/planned_debt match inside hyphenated code identifiers (HACK-LEVEL, PROBE-TODO) — debt rules are whole-text, not comment-anchored #2537, priority: high per the COBOL/JCL convention).
• Corpus: keyword-rosetta PR jcl transparency #10 (ENGINE_REF=pull/2622/head; cobol gate 75 assertions PASS,
  scheme 72 PASS; bias artifacts regenerated — the regen also folds in stale fix(detector): exclude synthetic slicer bucket names from orphan/duplicate census (#2547) #2618 cache drift,
  documented in the PR).
• Capstone: PR docs(language-status): cobol §10 rosetta consistency capstone (#2567 sweep) #2623 (docs/language_status/cobol.md §10).

Close criterion

Every remaining deviation is ledgered as intended morphology (args, globals) or blocked on a
named cross-cutting issue (state_mutation → #2546; high_risk_execution/io → #2535; §3 →
downstream). Per the epic, that satisfies the close condition — recommending close once #2622 and
rosetta PR #10 merge; closing is your call.

[squid-protocol]

Closing per the sweep summary above: 6🔴/4🟡 → 3🔴/4🟡, and every remaining deviation is ledgered as intended morphology (args → cobol-args-clause-match-morphology, globals → cobol-working-storage-globals) or blocked on a named cross-cutting issue (state_mutation → #2546, high_risk_execution/io → #2535, §3 → downstream re-baseline). Choreography landed: gitgalaxy#2622 + #2623 merged, keyword-rosetta#10 merged, KEYWORD_ROSETTA_REF → 89911a8, rosetta-audit green on main against the new pin.