workflow-fix: vectorize rule covers permutation/bootstrap draw loops

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #778 (emitting agent: orchestrator, /issue 778 session).

Goal

Extend .claude/rules/vectorize-many-cell-fits.md (and the planner/critic enforcement pointers) to explicitly cover many-DRAW closed-form statistical loops — permutation / bootstrap / null-draw batteries — not only gradient-descent fits.

Workflow gap

• Bug observed: #778's stage-two null battery ran a serial per-draw permutation loop (perm_null_draws: two full pool-mean passes over a 1783×28×3584 float64 pool PER DRAW) at ~4.1 s/draw — ~15h projected vs the plan §8 estimate of 1h — after the round raised n_draws 200→1000. The fix (batched subset-sum GEMM over all draws) is a ~70× win, exactly the rule's 50-100× class.
• Why it is a workflow gap: the rule's title, diagnostic signature, and plan-time enforcement (planner.md §9 compute-character carve-out, critic.md Methodology lens 10(iii)) are all framed around gradient-descent fits (MLP/AdamW/LOCO). A closed-form many-draw loop with identical overhead-bound economics does not pattern-match, so neither the planner nor the critic ensemble flagged the serial loop at plan time, and the CLAUDE.md always-on bullet's examples also omit draw loops.
• Confidence (emitter): high

Proposed change (candidate diff sketch — refine in planning)

.claude/rules/vectorize-many-cell-fits.md
+ frontmatter description: add "many-draw statistical loops (permutation /
+   bootstrap / null-draw batteries)" alongside GD fits
+ ## The diagnostic signature: add "a per-draw Python loop that re-reduces a
+   large fixed pool (means/sums/covariances) every draw — precompute the pool
+   reduction once and batch all draws as one GEMM (subset-sum identity)"
+ ## Worked incidents: add #778 (perm_null_draws, 4.1 s/draw, 5x draw raise,
+   ~70x vectorized win)
CLAUDE.md "Vectorize / parallelize compute by default" bullet
+ add "never a serial per-draw permutation/bootstrap loop over a large fixed
+   pool" to the example list

Scope / surfaces

• Primary target: .claude/rules/vectorize-many-cell-fits.md
• Secondary: CLAUDE.md (the always-on vectorize bullet's example list), .claude/agents/planner.md §9 / .claude/agents/critic.md Methodology lens 10(iii) if their wording needs the same widening.
• Grep before editing: grep -rln 'many-cell' .claude/ CLAUDE.md

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks passes; lessons index stays in sync.
• This session runs under the standard recursion-guard contract.

Provenance

• workflow_fix_target: .claude/rules/vectorize-many-cell-fits.md
• fingerprint: 0ab249f65d8d

Surfaced prose (verbatim): the orchestrator of /issue 778 observed the serial null-battery loop mid-round, measured 4.1 s/draw via py-spy, and routed the code fix through an experiment-implementer round; this task is the rule-file widening so the NEXT plan catches it at plan time.