control corpus Phase 5: pin keyword-rosetta as a cross-language consistency gate in gitgalaxy CI #2557

Successor to #1096 (the multi-language control corpus), which is closed as delivered by keyword-rosetta: 46/46 registry languages locked and gate-passing, with that repo's own CI running tools/verify_language.py on any changed data/ folder. What #1096's plan called Phase 5 — the gitgalaxy-side integration — is this issue.

Scope

1. Pinned-tag CI job in this repo, mirroring the crucible pin mechanism (LANGUAGE_CRUCIBLE_REF + tests/_crucible_pin.py): check out keyword-rosetta at a pinned tag, run its verifier across all 46 folders against the PR's engine build, fail on gate regressions. This turns the corpus from a point-in-time audit into a standing regression gate the same way the golden masters work.
2. Re-baseline protocol documented for intentional engine changes: each of the still-open corpus findings (prism/detector: string-literal shielding is selective by signal family (if/try/open count inside a literal, eval does not) #2535–fragile_debt/planned_debt match inside hyphenated code identifiers (HACK-LEVEL, PROBE-TODO) — debt rules are whole-text, not comment-anchored #2537, javascript func_start/args parse 'with (x) {}' as a function named 'with' #2539–orphan census counts non-function shapes on non-call-graph languages, making orphan→api conversion unauditable #2547, html func_start is unreachable end-to-end: the polyglot splitter consumes <script>/<style> tags before html's rules run #2549) will, when fixed, change planted-corpus expected values — the corpus already has the precedent (keyword-rosetta commit 0e8ebfa re-baselined embedded_python api 3→6 after PR fix(core): wire embedded_python into Python-family indentation-scoping #2548 landed here). The pin bump + expected-signals re-baseline should be one documented cross-repo step, like BUMPING_THE_CRUCIBLE_PIN.md.
3. Optional, carried from Feature: multi-language control corpus to validate structural extraction equivalence across languages #1096's original proposal and explicitly not delivered in Phases 1–4: wrapping each language folder in its idiomatic project/manifest shape to stress guidestar_lens.py's Intent Lock. Today only python (pyproject.toml), javascript (package.json) and rust (Cargo.toml) carry manifests; the other 43 are bare probe shells, which keeps planted counts exact but leaves the Intent Lock path untested by this corpus. Decide deliberately whether that's worth the count-perturbation risk before building it.

References

• Feature: multi-language control corpus to validate structural extraction equivalence across languages #1096 (closed epic) and its status comments for the phase history
• keyword-rosetta: SPEC.md, docs/GATING.md, deviation_ledger.json, bias report
• Crucible-pin precedent: docs/self_scan/BUMPING_THE_CRUCIBLE_PIN.md, tests/_crucible_pin.py

🤖 Generated with Claude Code

https://claude.ai/code/session_01M11wL1mSgk1fghxz1PAWGY