workflow-fix: verification gates must test the live dispatched path

Overview / Motivation

Auto-filed from the 2026-07-02 #779 unfixed-grid launch: the grid's --verify-vectorized flag ran vectorized_mlp_skill.assert_matches_reference() — a self-check of an UNUSED helper — while the live ridge path (the actual 17k-fit hot loop) had zero verification coverage. The session's launch note claimed "vectorized" on the strength of this green gate; the run was the unvectorized 18-20h path.

Goal

Rule: any verification/equivalence/vectorization gate flag must exercise the CODE PATH ACTUALLY DISPATCHED, not a sibling helper; a gate that passes without touching the hot loop is a hollow gate and a review blocker.

Workflow gap

• Bug observed: --verify-vectorized gated an unused helper; the orchestrator + plan treated its PASS as evidence the grid was vectorized (launch marker: "0-GPU vectorized").
• Why it is a workflow gap: no rule tells implementers/code-reviewers that verify-gates must cover the dispatched path; the code-review ensemble passed r6/r7 with the hollow gate in place.
• Confidence (emitter): high

Proposed change (refine in planning)

• .claude/rules/code-style.md: add a "verification gates test the live path" bullet (a --verify-X flag must assert on the exact function(s) the run dispatches; asserting on an unused sibling is a FAIL).
• .claude/agents/code-reviewer.md: one line in the review checklist — when a diff carries a verify/equivalence flag, confirm the gated function is the one the entrypoint calls.

Scope / surfaces

• Primary: .claude/rules/code-style.md, .claude/agents/code-reviewer.md.

Constraints / invariants

• ruff on touched files; keep the bullet short (the lessons-index stays lean).

Provenance

• workflow_fix_target: .claude/rules/code-style.md
• origin: PM-chat root-cause 2026-07-02 (#779 hollow --verify-vectorized gate)