workflow-fix: artifact-reuse (h) check for train-input mixes

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a <!-- workflow-fix-candidate v1 --> block raised on task #734 round 4 (emitting agent: experiment-implementer).

Goal

Add an (h) reuse-fitness check to .claude/rules/artifact-reuse.md covering reused training-input mixes / on-policy response caches — verify each reused mix file resolves on HF via list_repo_files AND is fetchable on the target backend BEFORE plan approval; if not fetchable, the plan must include a self-contained regen phase or upload the mix first.

Workflow gap

• Bug observed: A plan that reuses a parent experiment's TRAINING-INPUT MIX (a train/*.jsonl the parent built but never uploaded) passed planning + 3 review rounds, then crashed phase2 at the pre-train assert data_path.exists() on the git-clone-only GCP lane because the mix is on neither HF repo and the lane cannot stage it.
• Why it is a workflow gap: The (a)–(g) artifact-reuse fitness check verifies reused ADAPTERS resolve on HF (check (e)) but has NO check that a reused TRAINING-INPUT MIX / on-policy cache is fetchable on the TARGET BACKEND; the fact-checker verifies adapter paths but not train-mix paths.
• Confidence (emitter): medium

Proposed change (candidate diff sketch — refine in planning)

## The checklist
+ - **(h)** for reused TRAINING-INPUT artifacts (train/*.jsonl mixes, on-policy
+   response caches): the file resolves on HF via `list_repo_files` AND is
+   fetchable on the TARGET backend — the git-clone-only GCP lane stages NO
+   VM-local `data/`, so a mix the parent built but never uploaded is unreachable
+   there. If not fetchable: the plan includes a self-contained regen phase
+   (reuse the parent's deterministic build blocks) OR uploads the mix first.
+   (incident #734 round-4: #664's seed-42 marker mix was on neither HF repo;
+   phase2 crashed at the pre-train assert on the GCP lane.)

Plus parallel updates:

• .claude/agents/consistency-checker.md — extend the artifact-reuse section to call out the (h) check (the consistency-checker is the agent that runs the reuse-fitness pass).
• .claude/agents/planner.md — §11 hyperparameter grounding already requires Source: per knob; add an analogous "Source:" line per REUSED INPUT-DATA artifact (train mix, on-policy cache, eval JSON) that names how the file is fetched on the target backend.
• .claude/agents/critic.md Methodology lens item 9 (artifact reuse) — extend to verify (h) alongside (a)-(g).

Scope / surfaces

• Primary target: .claude/rules/artifact-reuse.md
• Sibling: .claude/agents/consistency-checker.md, .claude/agents/planner.md, .claude/agents/critic.md
• Grep the workflow surface for the pattern ((a)-(g), artifact-reuse, reuse-fitness) before editing — there may be other call sites that enumerate the (a)-(g) checks (code-reviewer.md, etc.) that need to grow a (h) reference too.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py 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: e6a24e31f459

target_file: .claude/rules/artifact-reuse.md bug_observed: A plan that reuses a parent experiment's TRAINING-INPUT MIX (a train/*.jsonl the parent built but never uploaded) passed planning + 3 review rounds, then crashed phase2 at the pre-train assert data_path.exists() on the git-clone-only GCP lane because the mix is on neither HF repo and the lane cannot stage it. why_workflow_gap: The (a)-(g) artifact-reuse fitness check verifies reused ADAPTERS resolve on HF (check (e)) but has NO check that a reused TRAINING-INPUT MIX / on-policy cache is fetchable on the TARGET BACKEND; the fact-checker verifies adapter paths but not train-mix paths. proposed_change: Add an (h) check to artifact-reuse.md — "training-input mixes / on-policy caches: verify each reused mix file resolves on HF via list_repo_files AND is fetchable on the target backend (the git-clone-only GCP lane cannot stage a VM-local-only mix) BEFORE plan approval; if not fetchable, the plan must include a self-contained regen step or upload it" — and have the fact-checker verify reused train-mix paths the same way it verifies adapter paths. diff_sketch: |

The checklist

• • (h) for reused TRAINING-INPUT artifacts (train/*.jsonl mixes, on-policy
• response caches): the file resolves on HF via list_repo_files AND is
• fetchable on the TARGET backend — the git-clone-only GCP lane stages NO
• VM-local data/, so a mix the parent built but never uploaded is unreachable
• there. If not fetchable: the plan includes a self-contained regen phase
• (reuse the parent's deterministic build blocks) OR uploads the mix first.
• (incident #734 round-4: #664's seed-42 marker mix was on neither HF repo;
• phase2 crashed at the pre-train assert on the GCP lane.) confidence: medium related_task: #734