workflow-fix: artifact-reuse check (l) — instrument validity-domain transfer

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix prose note raised on task #1417 (emitting agent: analyzer, first-pass interpretation).

Goal

Add a reuse-fitness checklist item to .claude/rules/artifact-reuse.md: a validity-domain transfer check — before reusing a fit/analysis INSTRUMENT (code artifact), verify the new data regime does not cross a validity boundary the instrument's own docs/comments declare (n ranges, dof caps, selection fallbacks / registered mitigations), and engage the declared mitigations when it does.

Workflow gap

• Bug observed: #1417 reused the frozen fit825 ridge instrument on judge-filtered row subsets (per-fold n_train < d=3584, the near-interpolation regime) without engaging the instrument's OWN documented mitigations (GCV_DOF_CAP = 0.9, lambda_selection="inner-group-cv", documented at scripts/issue825_fit_cells.py lines 68-91 from the #1335 incident). The GCV lambda selection collapsed on those subsets (held-out R² −0.6…−1.5 where supersets and matched-n subsamples fit at +0.3…+0.65), voiding the run's entire map-identity verdict layer — found only by the analyzer post-run.
• Why it is a workflow gap: the reuse fitness check (a)-(k) verified recipe identity, throughput (i), and parent lineage (k), but no item asks whether the NEW consumption regime crosses a validity boundary the reused instrument itself declares. Check (b) (valid measurement regime) is DV/question-scoped, not instrument-doc-scoped — it did not bind here.
• Confidence (emitter): high (the instrument's own docstring names the exact failure + mitigations; a checklist question would have caught it at plan time).
• verified-at-filing: grep -niE 'validity[- ]domain|validity boundary|dof cap' .claude/rules/artifact-reuse.md → 0 hits in the target (absence-of-guard claim; adjacent check (b) at line ~108 covers measurement-regime-for-the-question, not instrument-declared validity bounds — read in context, it does not implement this change); git log --oneline --since='7 days ago' -- .claude/rules/artifact-reuse.md → 5 commits, none touching validity-domain transfer (2026-07-18).

Proposed change (candidate diff sketch — refine in planning)

• New item under the reuse fitness checklist (sibling of (i) throughput):
• (l) validity-domain transfer (reused fit/analysis instruments): read the
• instrument's own docs/comments for DECLARED validity boundaries (n-vs-d
• regimes, dof caps, selection fallbacks, registered mitigations) and check
• the new consumption regime against them; crossing a declared boundary
• requires engaging the declared mitigation (or a stated justification) in
• the plan. A reused instrument consumed outside its self-declared domain is
• a REVISE (methodology-baselines lens / consistency-checker cross-check).
• Driving incident: #1417 x fit825 GCV collapse (#1335 mitigations unengaged).

Scope / surfaces

• Primary target: .claude/rules/artifact-reuse.md
• Grep the workflow surface for enforcement mirrors before editing (grep -rln 'artifact-reuse' .claude/agents/planner.md .claude/agents/consistency-checker.md .claude/rules/critic-lens-reference.md CLAUDE.md) and update the enforcement pointers (planner step 5 self-attestation, consistency-checker, Methodology lens item 9) if the item list is enumerated there; list every hit in the plan.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; if workflow.yaml or CLAUDE.md change, they stay consistent with the rule file.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).

Provenance

• workflow_fix_target: .claude/rules/artifact-reuse.md
• fingerprint: 836c079fd4fe

Verbatim surfaced prose (analyzer, #1417 first pass): ".claude/rules/artifact-reuse.md: the reuse fitness check (a)-(k) verified recipe identity but not validity-domain transfer — the frozen fit825 instrument's own docstring bounds its GCV default to the n_train > D regime, and #1417 moved it to filtered subsets (n_train < 3584) without engaging the documented mitigations. A checklist item 'does the new data regime cross a validity boundary the reused instrument's own docs/comments declare (n ranges, dof caps, selection fallbacks)?' would have caught this at plan time."