daily-fix: judge transport errors are RETRY, never dropped d

Overview / Motivation

Auto-filed by the /daily 2026-07-08 problem sweep (transcript-mined; emitting agent: /daily orchestrator).

Goal

Distinguish judge transport errors from content drops in the drop-never-coerce rule.

Workflow gap

• Bug observed: Transcript c16b10ca (issue-1090) 09:34:17Z interp-critique r1: the dropped judge draws were stored API-529 overload errors, freely re-judgeable; drop asymmetry 400/1000 trained draws on C2-icl; 10:07:34Z re-judge recovered 2,635/2,638 (zero refusals), all cells 200/200.
• Why it is a workflow gap: Rule 9 (drop-never-coerce) does not distinguish content-informative judge failures (REFUSAL/non-numeric/out-of-range) from transport failures; a pipeline persisting transport errors as drops silently censors arms, mimicking a selection artifact.
• Confidence (emitter): medium

Proposed change (candidate diff sketch — refine in planning)

Rules 9/23 amendment: transport errors (429/529/timeout/connection) are RETRY-with-backoff, never drops; the per-arm dropped-draw report splits content-drops vs transport-errors; a companion note that eval/batch_judge.py should not persist error:True entries counted as drops (library change = sibling infra task, src/ off-surface here).

Scope / surfaces

• Primary target: .claude/rules/llm-judging.md
• Grep the workflow surface for the pattern before editing and update every hit; list them in the plan.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks + --check-references pass; ruff on touched files passes.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

• workflow_fix_target: .claude/rules/llm-judging.md
• origin: transcript-mined by /daily 2026-07-08 problem sweep
• evidence: mine-1112-1090 P9 (c16b10ca 09:34:17Z, 10:07:34Z)