workflow-fix: judge caches must key on rubric+question+completion (llm-judging.md)

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a prose follow-up raised on task #810 (emitting agent: interpretation-critic).

Goal

Add an llm-judging.md guideline: any judge cache MUST key on (rubric/behavior id, question, completion), never completion content alone.

Workflow gap

• Bug observed: JudgeCache._hash_key (src/explore_persona_space/eval/batch_judge.py:107-110) omits the rubric/behavior from the cache key, so refusal-rubric judgments leaked into harmful_compliance E0 scores via the shared content-keyed cache dir on task #810 — 64% of high harmful_compliance scores were refusal-rubric judgments; 99.1% of flagged rows share their exact (reasoning, score) pair with the refusal file (vs 18.7% baseline), proving cache leak over a prompt bug.
• Why it is a workflow gap: .claude/rules/llm-judging.md is the binding recipe every judged-DV plan loads, and it is silent on cache keying — any future multi-behavior judged run that reuses completions across rubrics with a shared cache will silently corrupt scores the same way. The library fix itself is out of workflow-fix scope (src/ eval code); the rule-file guideline is in scope and is what makes planners/critics catch it.
• Confidence (emitter): high

Proposed change (candidate diff sketch — refine in planning)

+ In .claude/rules/llm-judging.md § G (Reproducibility / reporting) or a new
+ § C item: "Judge-cache keying: any judge result cache MUST include the
+ rubric/behavior identifier (plus question + completion) in its key —
+ NEVER completion content alone. A content-only key silently returns
+ another behavior's judgment for the same completion when completions are
+ reused across rubrics (incident #810: JudgeCache._hash_key omitted the
+ rubric; refusal judgments leaked into harmful_compliance E0, 64% of high
+ scores). Planners name the cache key fields; critics REVISE a shared
+ multi-rubric cache without a rubric-bearing key."

Scope / surfaces

• Primary target: .claude/rules/llm-judging.md
• Grep the workflow surface for the pattern before editing (grep -rln 'judge cache\|JudgeCache\|cache-dir\|cache_dir' .claude/ CLAUDE.md scripts/) and update every hit; list them in the plan. (Enforcement pointers in planner.md §6 / critic.md Statistics & Measurement lens may warrant a one-line cross-reference; the spawned planner decides.)

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/. The batch_judge.py library fix is a SEPARATE follow-up (out of scope here).
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; if workflow.yaml or CLAUDE.md change, they stay consistent with the rule file.
• 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/llm-judging.md
• fingerprint: 7589708532b4

Surfaced prose (interpretation-critic #810 round 1, verbatim): ".claude/rules/llm-judging.md should add a guideline that any judge cache MUST key on (rubric/behavior, question, completion), not content alone — the JudgeCache._hash_key gap will silently corrupt any multi-behavior judged run reusing completions across rubrics (the library fix itself is out of workflow-fix scope; the rule-file guideline is in scope)."