workflow-fix: vectorized_mlp_skill.py helper referenced but missing

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from an orchestrator-observed gap (confirmed empirically while locating the vectorized fit code for a #658 re-analysis).

Goal

Fix the phantom canonical-helper reference: .claude/rules/vectorize-many-cell-fits.md and CLAUDE.md both cite src/explore_persona_space/analysis/vectorized_mlp_skill.py as the canonical vectorized many-cell-fit helper, but that file does not exist in the repo. Either create the helper or repoint the references to the actual vectorized code.

Workflow gap

• Bug observed: git ls-files | grep -i vectoriz returns ONLY .claude/rules/vectorize-many-cell-fits.md — there is NO vectorized_mlp_skill.py anywhere in the tree. Yet the rule (lines ~62, ~83) and the CLAUDE.md "vectorize-many-cell-fits" / compute-character carve-out bullet both name src/explore_persona_space/analysis/vectorized_mlp_skill.py as "the canonical helper." An agent instructed to "use the canonical helper" cannot find it.
• Why it is a workflow gap: the rule + CLAUDE.md are workflow surface; they point at a non-existent file, so the guidance is unfollowable as written. The actual vectorized many-cell MLP fit currently lives inline in per-issue scripts (e.g. _fit_mlp_ensemble_loco, the torch.vmap ensemble, in scripts/issue658_fit_predictors.py), not in a shared helper.
• Confidence (emitter): high (file absence is verified by git ls-files)

Proposed change (candidate diff sketch — refine in planning)

Planner picks ONE:

• (a) CREATE src/explore_persona_space/analysis/vectorized_mlp_skill.py as the reusable helper, factoring the vmap-ensemble LOCO MLP out of scripts/issue658_fit_predictors.py::_fit_mlp_ensemble_loco (the existing, exactness-tested implementation), and have the rule point at it; OR
• (b) REPOINT the rule + CLAUDE.md references to the actual in-repo vectorized code (scripts/issue658_fit_predictors.py::_fit_mlp_ensemble_loco) and drop the "canonical helper at " framing until a shared helper is genuinely extracted.

Option (a) is the better long-term fix (the rule's intent is a REUSABLE helper); (b) is the honest stopgap. Either removes the dangling reference.

Scope / surfaces

• .claude/rules/vectorize-many-cell-fits.md
• CLAUDE.md (the vectorize-many-cell-fits / compute-character carve-out bullet under "CPU-only phases don't hold GPU pods")
• (if option a) src/explore_persona_space/analysis/vectorized_mlp_skill.py (NEW) — note this path is OUTSIDE the normal workflow surface, but creating the file the rule mandates is in-scope for closing the gap; the planner decides (a) vs (b).
• Grep first: grep -rln 'vectorized_mlp_skill' .claude/ CLAUDE.md to catch every reference.

Constraints / invariants

• Keep the rule's paths: frontmatter and CLAUDE.md wording consistent with whichever option is chosen.
• If option (a): the extracted helper must preserve _fit_mlp_ensemble_loco's exactness contract (the _assert_mlp_exactness oracle) — do not silently change fit numerics.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

• workflow_fix_target: .claude/rules/vectorize-many-cell-fits.md
• fingerprint: 6e8565604a68

target_file: .claude/rules/vectorize-many-cell-fits.md, CLAUDE.md bug_observed: git ls-files shows no vectorized_mlp_skill.py anywhere in the repo, yet .claude/rules/vectorize-many-cell-fits.md and CLAUDE.md both name it as the canonical helper, so an agent told to use the canonical helper cannot find it. why_workflow_gap: the rule + CLAUDE.md (both workflow surface) point at a non-existent file; the actual vectorized many-cell MLP fit lives inline as _fit_mlp_ensemble_loco in scripts/issue658_fit_predictors.py. proposed_change: Either create src/explore_persona_space/analysis/vectorized_mlp_skill.py by factoring out _fit_mlp_ensemble_loco, or repoint the rule + CLAUDE.md references to the actual in-repo vectorized code; remove the dangling reference either way. diff_sketch: |

• src/explore_persona_space/analysis/vectorized_mlp_skill.py — the reusable

• (option b) the canonical vectorized many-cell fit lives in
• scripts/issue658_fit_predictors.py::_fit_mlp_ensemble_loco (torch.vmap
• ensemble, exactness-tested); extract to a shared helper when reused again. confidence: high related_task: #658