workflow-fix: wire the both-arms mapping rule into Statistics lens item 15

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a gap the orchestrator observed during interactive planning of a context→answer mapping issue (2026-08-05). The user asked what the standing both-arms mapping rule says exactly; reading it surfaced that its own stated enforcement does not exist in the files it names.

Goal

Add the prefix-arm/context-arm both-arms requirement as a clause of the existing representation-mapping registration — Statistics lens item 15 + the planner §6 registration block + statistics-critic item 15 + experiment-guidelines guideline 11 — using the same trigger and the same stated-deviation escape as its identity+bias / kNN / pooling-convention siblings.

Workflow gap

• Bug observed: the CLAUDE.md standing rule "Prefix mapping AND context mapping — run BOTH in every experiment" ends with "The planner names both arms in §4 Design; the critic REVISEs a plan that silently drops one without a stated exemption." Neither clause is implemented: the mapping arms appear nowhere in planner.md, critic.md, critic-lens-reference.md, statistics-critic.md, or experiment-guidelines.md. The rule is carried only by the always-on CLAUDE.md bullet and by .claude/skills/issue/SKILL.md's inline-round duty.
• Why it is a workflow gap: three structurally identical representation-mapping disclosures ARE wired into those exact surfaces — the identity+bias baseline and the kNN-retrieval read (#1604, merge 4c4e79688b) and the pooling-convention row (#1746, merge 4238694ae9) — all under Statistics lens item 15, which fires on precisely the trigger this rule needs ("If the plan FITS a map between activation summaries (context→answer, prefix→context, cross-model / cross-framing reparameterization — any v_X→v_Y predictor)"). The both-arms rule is the fourth sibling of that set and is the only one absent, so a critic working item 15 top-to-bottom will check pooling and the two baselines and never check the arms. Two shipped one-arm rounds are on record and both were caught by the user rather than the pipeline: #958 (2026-07-04, a one-arm capture, which is why the routing-section capture-time clause was added) and #779's 2026-07-14 inline pre-image round (context-only).
• Confidence (emitter): medium. There is a competing reading under which the CLAUDE.md sentence is a normative instruction to agents that all load CLAUDE.md, so the rule is "enforced, weakly" rather than unwired. The planner should adjudicate that with the files open; the asymmetry against the three wired siblings is the substantive argument, and the two user-caught incidents are the evidence that the weak form is insufficient in practice. A reasoned no-change report is an acceptable outcome.
• verified-at-filing: grep -ciE "prefix-based|prefix arm|both mapping arms|prefix mapping AND context" <target> → 0 hits in all four target files, and 0 in planner.md / critic.md as well (2026-08-05). Absence-of-guard claim, so the 0-hit in-target result is the evidence. Correct-insertion-point confirmed positively in the same pass: grep -c "identity_bias_predict|identity+bias|identity-family" → critic-lens-reference.md 5, planner-section-reference.md 4, experiment-guidelines.md 2, statistics-critic.md 1. Landed-fix check: git log --oneline --since='14 days ago' on the three named CLAUDE.md targets returns 4238694ae9 (#1746 pooling-convention row, 7 wired surfaces) — inspected, it wires POOLING, not the mapping arms, so it is a sibling landing and not this fix.

Proposed change (candidate diff sketch — refine in planning)

critic-lens-reference.md, Statistics lens item 15 — append a third clause:
+   Additionally verify the plan runs BOTH mapping arms — prefix-based (prefix =
+   everything before the user query) AND context-based (context = prefix + query) —
+   as paired arms of the same design, or names the omission as an explicit stated
+   deviation carried into the clean-result as a scope caveat. REVISE a mapping plan
+   that silently runs one arm. Not a REVISE when the plan fits no map, or the
+   one-arm scope is stated + justified (e.g. a substrate whose prefix is the
+   constant chat template, making the prefix arm a degenerate constant-input floor).
+   Full rule: CLAUDE.md § "Prefix mapping AND context mapping".

planner-section-reference.md §6 — mirror as a "Mapping-arms row (same trigger)" block
statistics-critic.md item 15 — mirror the critic clause
experiment-guidelines.md guideline 11 — extend the title + body to name the arms

Scope / surfaces

• Primary targets: .claude/rules/critic-lens-reference.md, .claude/rules/planner-section-reference.md, .claude/agents/statistics-critic.md, .claude/rules/experiment-guidelines.md
• These are exactly the four surfaces #1604 and #1746 wired for the sibling representation-mapping disclosures; follow their landed shape rather than inventing a new location.
• If CLAUDE.md's bullet is updated to record the wiring (as the identity+bias bullet does), keep it consistent with the rule file.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py (no flags) passes; ruff on touched files passes.
• The "unless explicitly stated otherwise" escape is load-bearing and must survive: this fix makes the deviation checked, never forbidden.
• Do not touch the SKILL.md inline-round duty or the routing-section capture-time clause — both already carry the rule correctly.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).

Provenance

• workflow_fix_target: .claude/rules/critic-lens-reference.md,.claude/rules/planner-section-reference.md,.claude/agents/statistics-critic.md,.claude/rules/experiment-guidelines.md
• fingerprint: a8a502c50011

Surfaced by the orchestrator during interactive planning (no subagent involved). The user asked "what does the project wide rule say exactly?" about the both-arms mapping rule; quoting it surfaced that its closing enforcement sentence names files that do not implement it.