rosetta[shell]: cross-language consistency — 1 red / 5 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)

In band — no structure-level deviations.

2. Signal / keyword deviations

• 🟡 globals: 3 vs median 2 (+50%)
• 🟡 io: 4 vs median 3 (+33%)
• 🟡 keyword_hits: 169 vs median 247.5 (-32%)
• 🟡 safety_bypasses: 3 vs median 2 (+50%)

3. Risk-exposure consequences (downstream)

• 🔴 risk_cognitive_load: 29.5468 vs median 15.1444 (+95%)
• 🟡 risk_documentation: 74.608 vs median 57.1208 (+31%)

Ledgered shapes touching shell

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
shell-test-brackets-in-branch	engine-semantic	—	(notes)	shell's branch rule counts the test brackets [ and ] as branch tokens (finditer on main.sh: if,[,],then,fi = 5 for one conditional) — every bracketed test costs +2 branc…
shell-trap-signal-env-collision	keyword-overlap	—	(notes)	'trap : TERM' counts globals: the TERM signal name collides with the TERM env-var keyword in shell's globals rule (a.sh globals 3 for 2 planted)
orphan-conversion-opaque-counts	upstream-question	#2547	(notes)	On languages without a same-file call graph the orphan census includes non-function shapes, so the orphan->api conversion count diverges from visible function counts: sh…
string-literal-selective-shielding	upstream-question	#2535	b.sh	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.sh · never needs root and the disk words stay in prose (comment): stripped: contributed 0
• b.sh · note="if eval fails, try curl" (string): if (branch +1), curl (io +1) AND eval (safety_bypasses +1) all COUNT — eval is a shell BYPASS keyword, not high-risk-family, and it is not shielded here, supporting the theory that only the high-risk family gets literal shielding in the python/js profile. Phantom branch context triples state: counter=/note= 2 raw -> 6. Ledger: string-literal-selective-shielding, state-flux-branch-weighting
• c.sh · : "$hack_level" (code): inert (underscore + lowercase)

Reproduce / evidence

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