workflow-fix: codex twin round-bound 1-3 vs cap-5 delta rounds

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a prose follow-up surfaced by codex-clean-result-critic (composer) on task #952 r4.

Goal

Relax the revision_round bound in .claude/agents/codex-clean-result-critic.md + .claude/agents/codex-interpretation-critic.md from 1-3 to 1-5, with rounds 4-5 valid only as delta-scoped re-reviews after a reconciler-bound REVISE; refuse only malformed rounds (≤0, >5, non-integer).

Workflow gap

• Bug observed: The #952 r4 delta dispatch — a legitimate post-reconciler round under the cap-5 ensemble policy (workflow.yaml § ensemble_review round cap 5) — violated the twin specs' stated 1-3 round bound ("epm:failure on anything else"); the composer had to proceed on a self-declared stale-spec assumption.
• Why it is a workflow gap: the twin specs' round bound predates the all-rounds/cap-5 policy; any future round-4/5 delta review re-hits the contradiction, and a literal-minded composer would falsely post epm:failure on a valid dispatch.
• Confidence (emitter): medium

Proposed change (candidate diff sketch — refine in planning)

- rounds 1-3 only; post epm:failure on anything else
+ rounds 1-5; rounds 4-5 are valid only as delta-scoped re-reviews following a
+ reconciler-bound REVISE (workflow.yaml round cap 5); refuse only malformed
+ rounds (<=0, >5, non-integer)

Scope / surfaces

• Primary target: .claude/agents/codex-clean-result-critic.md, .claude/agents/codex-interpretation-critic.md
• Grep first: grep -rln 'revision_round' .claude/agents/codex-*.md and align every twin with the same stale bound.

Constraints / invariants

• Workflow-surface only; keep the delta-scoped framing consistent with workflow.yaml § ensemble_review.

Provenance

• workflow_fix_target: .claude/agents/codex-clean-result-critic.md, .claude/agents/codex-interpretation-critic.md
• fingerprint: 88eb7e26dc46