rosetta[python]: cross-language consistency — 5 red / 3 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

• 🔴 comment_lines: 52 vs median 30 (+73%)
• 🔴 doc: 2 vs median 1 (+100%)
• 🔴 io: 6 vs median 3 (+100%)
• 🔴 state_mutation: 2 vs median 6 (-67%)
• 🟡 safety: 3 vs median 2 (+50%)
• 🟡 test: 3 vs median 2 (+50%)

3. Risk-exposure consequences (downstream)

• 🔴 risk_cognitive_load: 4.821 vs median 15.1444 (-68%)
• 🟡 risk_api_exposure: 7.35003 vs median 5.31104 (+38%)

Ledgered shapes touching python

defect	type	issue	evidence	summary
api-contextual-baseline-fix	engine-semantic	#2536	—	galaxyscope.py ~2145 ('THE CONTEXTUAL BASELINE FIX'): a file imported by others (popularity > 0) has orphaned_logic converted into api before recording
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
assert-overlaps-safety-and-test	keyword-overlap	—	—	python's test rule includes \bassert\b and the safety rule includes assert: one planted assert counts both
os-sys-prefix-overlaps-io	keyword-overlap	—	—	python-family io rule matches the os./sys
string-literal-selective-shielding	upstream-question	#2535	—	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.py · never calls eval and no while loop (comment): stripped: contributed 0 (prism comment stripping works)
• a.py · could open a socket in prose only (comment): stripped: contributed 0
• a.py · MESSAGE = "if eval fails, try open" (string): PARTIAL: if/try/open COUNT from inside the literal (branch +2, safety +1, io +1, baked into a.py expectations) but eval does NOT (feeds sec_tainted_injection instead) — literal shielding is selective by signal family; candidate upstream issue
• b.py · the try keyword stays in prose (comment): stripped: contributed 0
• c.py · HACK_LEVEL = level (code): reverse decoy works: identifier did not count, the # HACK comment did (fragile_debt = 1 exactly)

Reproduce / evidence

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

[squid-protocol]

Sweep update (2026-08-31)

Worked via the rosetta-language-sweep skill's five-cause taxonomy. Live numbers before starting
(tools/language_deviations.py python) didn't quite match this issue's original body — comment_lines
was red live but not listed in §1 above; treated the live numbers as authoritative per the skill's
guidance.

Before → after: 5🔴 / 3🟡 → 4🔴 / 3🟡 (37 comparable metrics).

Fixed (bucket 1: real engine bug)

• io: 6 → 4 (median 3). os\./sys\. matched ANY attribute access, overlapping globals'
  own os.environ/sys.argv/sys.path — a planted globals read was double-counted as io too.
  Negative lookaheads now carve out exactly those three tokens.
• test: 3 → 2 (median 2, now in-band). \bassert\b was double-counted by both safety and
  test — a bare assert is a validation signal, not a testing-framework signal. Removed from test.
• Engine PR: gitgalaxy#2626 (merged pending
  review). Corpus re-baseline: keyword-rosetta#13.
  Ledger: os-sys-prefix-overlaps-io (resolved for python), assert-overlaps-safety-and-test
  (python removed from languages_seen, still reproduces for c/perl).

Not fixed — bucket 5 (median inflation elsewhere, python is honest)

• safety: 3 vs median 2 (+50%) — matches planted intent (3) exactly.
• state_mutation: 2 vs median 6 (-67%) — matches planted intent (2) exactly; the median is
  inflated by the ×3 flux weighting in other languages (#2546).
  No python action; re-baselines when state-flux ×3 branch-context weighting: undocumented, and phantom-triggered by branch keywords inside string literals #2546 lands.

Not fixed — out of scope for this pass, filed separately

• doc: 2 vs median 1 (+100%) — the """/''' docstring pattern counts both the opening and
  closing delimiter as separate matches. Related to the string-literal-shielding upstream question
  (#2535); not re-investigated here.
• comment_lines: 52 vs median 30 (+73%), new finding, not python-specific: filed as
  #2625 — comment_lines = total_loc − coding_loc,
  but total_loc includes blank lines while coding_loc doesn't, so blank lines get folded into
  "documentation" engine-wide. Python's corpus shells just have a higher blank-line ratio than most
  language shells, so it surfaces loudest here (jcl's own §10 capstone independently noted the same
  root cause for its comment_lines residual).
• risk_api_exposure/risk_cognitive_load (§3): downstream shadows of the above; not chased
  individually per the epic's rule.

Remaining blocked on

• safety/state_mutation: cross-cutting state-flux ×3 branch-context weighting: undocumented, and phantom-triggered by branch keywords inside string literals #2546.
• doc: string-literal shielding question prism/detector: string-literal shielding is selective by signal family (if/try/open count inside a literal, eval does not) #2535 (if that lands, re-measure before assuming doc
  is still a real bug).
• comment_lines: new cross-cutting comment_lines conflates blank lines with real documentation (engine-wide, found via #2593) #2625 (engine-wide total_loc/coding_loc blank-line fix).

Closing criterion: not met yet — io is now amber (not green), safety/doc/comment_lines
remain out of band, all blocked on cross-cutting issues rather than anything left to do in python's
own rule file. Leaving this issue open pending #2546/#2535/#2625; re-run
tools/language_deviations.py python after any of those land to see if python re-baselines into
band on its own.

[squid-protocol]

Choreography completed (2026-09-01)

Follow-up to the sweep update above — the cross-repo loose ends are now closed:

• rosetta#13 closed, not merged: it was a stale duplicate of the already-merged Fix cascade delete hang by adding indexes to class and function tables #12 from a
  rewritten branch; its only remaining delta vs main would have reverted rosetta#11's skill
  update.
• rosetta#14 merged: Fix cascade delete hang by adding indexes to class and function tables #12 had landed with ENGINE_REF still at pull/2626/head (the
  restore-to-main step was skipped). chore(deps): bump actions/checkout from 4 to 6 #14 restores it — and fixes the CI gap it exposed:
  verify.yml's path filter didn't include ENGINE_REF, so a pin-only PR ran zero gates.
  It now triggers the full 46-gate sweep (which chore(deps): bump actions/checkout from 4 to 6 #14 then passed against engine main).
• KEYWORD_ROSETTA_REF bumped to 657ed07 (rosetta main post-chore(deps): bump actions/checkout from 4 to 6 #14); rosetta-audit
  dispatched on main against the new pin as the end-to-end proof.
• Capstone merged: PR docs(language-status): python §11 rosetta consistency capstone (#2593 sweep) #2630 (docs/language_status/python.md §11).

Issue state unchanged per the close-criterion note above: stays open, blocked on the
cross-cutting fixes (#2546 safety/state_mutation medians, #2535 doc, #2625 comment_lines) —
re-run tools/language_deviations.py python after any of those land.