daily-fix: reuse check (l) needs a call-shape bind; renames

Overview / Motivation

Auto-filed by the /daily 2026-07-26 problem sweep (route 2). Surfaced by 1 independent miner group(s) over the 2026-07-26 session transcripts.

Goal

Extend the artifact-reuse fitness check (l) with a CALL-SHAPE bind of the actual kwargs the new caller will pass, and add a symbol-rename whole-tree grep requirement to the crash-fix-rounds rule.

Workflow gap

• Bug observed: two reuse/refactor validity gaps each crashed a live run — a reused parent fit helper whose signature accepts a lambdas= kwarg but whose code path carries a runtime assert lambdas is None, and a class rename that fixed one call site and left the sibling script importing the old name.
• Why it is a workflow gap: .claude/rules/artifact-reuse.md binds kwarg NAMES against a signature (inspect.signature(...).parameters) but never binds the actual call, and .claude/rules/crash-fix-rounds.md has no whole-tree grep duty for a renamed symbol.
• Confidence (emitter): high
• verified-at-filing: grep -c -i 'call-shape\|runtime assert' .claude/rules/artifact-reuse.md → 0 (absence-of-guard evidence); context binding of the nearest existing surface — .claude/rules/artifact-reuse.md L697–718 "Signature smoke per kwarg the dispatcher passes" prescribes missing = dispatcher_kwargs - {f.name for f in fields(TrainLoraConfig)} and, for non-dataclass callees, inspect.signature(<fn>).parameters — a NAME-membership test that a signature accepting lambdas= passes by construction; check (l) at L484–513 prescribes reading the instrument's "docs/comments/module constants" for declared boundaries, not binding the call; grep -c -iE 'rename.*grep|grep -rn .<old' .claude/rules/crash-fix-rounds.md → 0, and the only two rename hits (L366, L417) concern renamed PATHS under a glob, not renamed symbols; git log --oneline --since='7 days ago' -- .claude/rules/artifact-reuse.md .claude/rules/crash-fix-rounds.md → 4 commits, none landing either duty (2026-07-26)

Evidence

• (i) Call-shape gap. Session 5c5a89e8, 2026-07-26T19:42:33Z: scripts/issue1689_fit_cells.py called the reused issue825_fit_cells.heldout_r2_sweep with a custom lambdas= grid. The parent's _ridge_predict_cached carries a runtime guard on the inner-group-cv plus inner-cache path: "issue825_fit_cells.py\", line 329, in _ridge_predict_cached\n assert lambdas is None, (\nAssertionError: inner-group-cv lambda selection scans the module LAMBDAS grid; a custom lambdas= grid is unsupported with an inner cache". The signature accepts the kwarg; the code path rejects it.
• Nothing at plan time and nothing in rounds R1 through R8 bound the actual call against that runtime guard, so the crash surfaced only after Phase B (30/30 on-policy pairs, roughly 13 hours) and Phase C (capture) had completed. Measured cost: crash-fix round R12 — implementer plus code review plus relaunch, roughly 13 minutes — plus the risk of losing 13 hours of upstream compute had the store not persisted.
• The existing instrument is name-scoped by construction: the L697 signature smoke asserts kwarg names are present in the callee's signature. lambdas IS present in heldout_r2_sweep's signature, so the smoke passes and the runtime assert still fires. Check (l) is doc-scoped: it directs the planner to read module comments and constants (its own worked example cites GCV_DOF_CAP and lambda_selection="inner-group-cv" at issue825_fit_cells lines ~66–91), which does not reach an assert buried at line 329 of a code path.
• (ii) Rename gap. Session 5c5a89e8, 2026-07-26T06:35:50Z: an earlier round (R5) renamed DispatchCall to DispatchItem in scripts/issue1689_haiku_u2_gen.py but not in scripts/issue1689_gen_onpolicy.py:297. The relaunched workload reached Phase B and died at import, taking the vLLM engine core with it: "line 297, in generate_and_filter\n from explore_persona_space.llm.api_dispatch import ( # noqa: E402\nImportError: cannot import name 'DispatchCall' from 'explore_persona_space.llm.api_dispatch'\nERROR 07-26 06:31:08 [core_client.py:564] Engine core proc EngineCore_DP0 died unexpectedly".
• The implementer's own failure lesson, emitted generalizes: yes, names the exact rule: "when fixing a class-level API rename, grep the whole scripts/ tree for the old name in the same round and fix every hit — sibling scripts drift until the next phase invokes them." Measured cost: one full crash-fix round (R9 implementer plus code review plus relaunch), roughly 20 minutes, plus a wasted pod launch cycle.

Proposed change

• In .claude/rules/artifact-reuse.md check (l) (L484–513), add a CALL-SHAPE BIND clause: for every reused fit/analysis helper, the plan records a probe of the ACTUAL call the new code will make — the exact kwargs at their exact values (or minimal stand-ins), executed at smoke shape — not a signature-name membership test. A parent helper can carry runtime asserts that reject a legal-looking kwarg combination.
• Name the diagnostic explicitly: a kwarg present in the signature is NOT evidence the call path accepts it. Direct the probe at the callee's body — grep -n 'assert\|raise NotImplementedError\|raise ValueError' <reused helper> for guards naming the kwargs the new caller passes — as the cheap companion to executing the call.
• Amend the L697–718 "Signature smoke per kwarg" step so it states its own limit: name-membership only; a runtime-guard rejection is out of its scope and belongs to the check (l) call-shape bind.
• Mirror the call-shape bind as a plan-time item in .claude/agents/planner.md §5, next to the existing reuse-fitness self-attestation, so the probe is recorded before compute is spent upstream of the fit.
• In .claude/rules/crash-fix-rounds.md, add a symbol-rename whole-tree grep duty: any round that renames a class, function, dataclass, or module-level constant runs grep -rn '<old_name>' scripts/ src/ in the SAME round and either fixes or explicitly dispositions every hit; the grep command and its hit disposition are recorded in the implementer marker.
• Have the code-reviewer verify the recorded grep is present whenever the diff contains a rename, so the duty is checked rather than self-attested.

Scope / surfaces

• Primary target: .claude/rules/artifact-reuse.md
• .claude/rules/crash-fix-rounds.md (the symbol-rename whole-tree grep duty)
• .claude/agents/planner.md (§5 plan-time mirror of the call-shape bind)
• .claude/agents/code-reviewer.md (verify the recorded rename grep)

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• uv run python scripts/workflow_lint.py passes (no-flags); ruff clean on touched files.
• This session runs under a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

• fingerprint: 55dc99aada75

• workflow_fix_target: .claude/rules/artifact-reuse.md

• fingerprint: PENDING

/daily 2026-07-26 route-2 filing. Miner refs: A-P2, A-P3.