workflow-fix: artifact-reuse fitness check gains throughput item (i)

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a user-approved transcript audit (2026-07-02 PM-chat session; user: "parallelization/vectorization is not properly being used ... Make all seven changes"). Emitting agent: orchestrator, via 5 parallel transcript-mining subagents.

Goal

Add a throughput fitness item (i) to the artifact-reuse checklist: reused fit/analysis code gets its inner loop and device routing inspected against vectorize-many-cell-fits.md, with fixes landing at the source module rather than the caller.

Workflow gap

• Bug observed: four incidents in one week came from reusing a parent task's fit/analysis code whose serial inner loop or hardcoded device pin nobody re-inspected: #761 reused #658's issue658_fit_predictors._ridge_predict_loco (Python loop over 50 LOCO folds, eigh+solve per fold) → 100x over plan (0.3h → 30h/behavior projected); #667's inline analysis reused #722's serial per-cell fit_cell (300-epoch MLP per cell × 84 cells, MLP already known-uninformative); #763 and #812 inherited the hardcoded DEVICE = "cpu" module constant from issue658_fit_predictors.py:143, which defeated vectorization even after ops were batched (#763: ~67 min/behavior on CPU vs ~2h total on GPU after the EPM_FIT_DEVICE=cuda cutover).
• Why it is a workflow gap: .claude/rules/artifact-reuse.md mandates the (a)-(h) fitness check for reused trained artifacts and data, and CLAUDE.md's "Reuse existing experiment code" bullet makes code reuse the default — but NO fitness item covers throughput character. Reuse currently propagates serial loops and stale device pins silently; each downstream task rediscovers the same blowup.
• Confidence (emitter): high.

Proposed change (candidate diff sketch — refine in planning)

In .claude/rules/artifact-reuse.md, extend the fitness checklist:

+ (i) Throughput fitness (reused CODE only — fit/analysis/eval helpers): before
+   adopting a parent's fit/analysis helper, READ its inner loop and device
+   routing. Two checks: (1) the per-cell/per-fold/per-draw axis is batched
+   (vectorize-many-cell-fits signature — a serial Python loop over cells,
+   folds, draws, or rows fails); (2) the device is parametrized (a hardcoded
+   DEVICE constant or implicit CPU pin fails — #763/#812 inherited
+   issue658_fit_predictors.py's DEVICE="cpu" and burned hours batched-on-CPU).
+   On a failure: fix at the SOURCE module (parametrize/batch it there) so
+   every future reuser inherits the fix — never a caller-side workaround.
+   State the inspection result (function name, batched-or-serial, device
+   handling) in the plan's reuse map / §9 row.

Mirror a one-line pointer in the CLAUDE.md artifact-reuse bullet ("(a)-(h)" → "(a)-(i)") and every workflow-surface file that enumerates the (a)-(h) list (grep for "(a)-(h)").

Scope / surfaces

• Primary target: .claude/rules/artifact-reuse.md
• Secondary: every file enumerating the fitness list — grep -rln '(a)-(h)' .claude/ CLAUDE.md and update each hit to (a)-(i).
• Cross-link: open task #869 Fix D adds a named-helper-actually-called review check (code-reviewer/consistency-checker side); this task owns the reuse-time inspection. Companion plain-infra task (filed in the same batch) parametrizes the concrete issue658_fit_predictors.py:143 DEVICE pin.

Constraints / invariants

• Workflow-surface only — the concrete DEVICE code fix belongs to the companion infra task, not this one.
• scripts/workflow_lint.py default run passes; CLAUDE.md and the rule file stay consistent.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 / a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

• workflow_fix_target: .claude/rules/artifact-reuse.md
• fingerprint: 0102a775d4d7

target_file: .claude/rules/artifact-reuse.md bug_observed: four incidents (#761 100x, #667, #763, #812) reused a parent's fit/analysis code whose serial inner loop or hardcoded DEVICE=cpu nobody re-inspected; the artifact-reuse fitness check (a)-(h) has no throughput item why_workflow_gap: code reuse is the project default but the reuse fitness checklist never inspects throughput character, so serial loops and stale device pins propagate silently to every downstream task proposed_change: add throughput fitness item (i) to the reuse checklist: reused fit/analysis code gets its inner loop and device routing inspected against vectorize-many-cell-fits, with fixes landing at the source module rather than the caller diff_sketch: |

• (i) Throughput fitness (reused code): inner loop batched? device

• parametrized? On failure fix at the SOURCE module; record the
  

• inspection (fn name, batched-or-serial, device) in the plan reuse map.
  

confidence: high related_task: n/a (user-chat transcript audit, 2026-07-02)