workflow-fix: Relocate live pods_ephemeral.json out of git t

Overview / Motivation

Auto-filed by the /daily Step-C parked-candidate sweep from a candidate parked on task #821.

Goal

Protect the live pods_ephemeral.json metadata from destructive repo-root git ops by relocating it out of the working tree with atomic writes, mirroring the #821 pods.conf fix.

Workflow gap

• Bug observed: scripts/pods_ephemeral.json is the same tracked-file-with-uncommitted-live-mutations class that wiped pods.conf: pod_lifecycle.py plain-writes it in-tree (EPHEMERAL_STATE = _PODS_EPHEMERAL_JSON_MAIN at :104; non-atomic EPHEMERAL_STATE.write_text at :276), so a repo-root destructive git op silently rewinds pod metadata (issue mapping, manual_override flags).
• Why it is a workflow gap: the failure originates in the workflow surface named below, not in any one experiment.
• Confidence (emitter): see parked note

Proposed change (candidate diff sketch — refine in planning)

• relocate live pods_ephemeral.json to /eps/ (mirror
• pod_config's #821 pods.conf contract: seed from tracked file, atomic
• tmp+os.replace write, never-drop guard for entries with a RUNNING pod)

• EPHEMERAL_STATE.write_text(json.dumps(payload, indent=2) + "\n")

• atomic write via tmp + os.replace; readers resolve the relocated path first

Scope / surfaces

• Primary target: scripts/pod_lifecycle.py
• Grep the workflow surface for the pattern before editing (grep -rln '<pattern>' .claude/ CLAUDE.md scripts/) 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 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: scripts/pod_lifecycle.py
• origin: parked candidate on task #821 at 2026-07-02T02:47:47Z

Verbatim parked note:

routed: parked: EPM_WORKFLOW_FIX_SESSION (recursion guard — logged, not auto-filed; see .claude/rules/workflow-fix-on-bug.md § Recursion guard). TWO candidates surfaced this session:

[1] source: prose-followup (implementer r1 report + plan §11)
target_file: scripts/pod_lifecycle.py
bug_observed: scripts/pods_ephemeral.json is the same tracked-file-with-uncommitted-live-mutations class that wiped pods.conf (EPHEMERAL_STATE plain write_text at pod_lifecycle.py:274); a repo-root destructive git op silently rewinds pod metadata (issue mapping, manual_override flags)
why_workflow_gap: live mutable state stored as uncommitted edits of a git-tracked file; no relocation/guard/self-heal analogue shipped for it in #821 (explicit non-goal, plan §11)
proposed_change: apply the #821 pattern to pods_ephemeral.json — relocate live copy to <git-common-dir>/eps/pods_ephemeral.json with seed fallback + atomic write; no --refresh-from-api analogue exists so scope a re-inference strategy
confidence: high

[2] source: orchestrator observation (this session)
target_file: .claude/agents/planner.md, .claude/agents/critic.md (grep 'model: claude-fable-5' across .claude/agents/*.md for the full hit set)
bug_observed: three consecutive planner spawns died with 'Autocompact is thrashing' (total_tokens: 0, ~8-11 tool uses) under the frontmatter pin model: claude-fable-5 + effort: xhigh; the same briefs succeeded immediately with a per-spawn model override to opus
why_workflow_gap: the pinned model's context window is too small for the always-on project context (CLAUDE.md+rules ~100KB) + a 93KB agent spec + tool outputs, so any planner/critic spawn without an override thrashes; orchestrators not knowing the workaround lose whole planning rounds
proposed_change: re-pin the affected agents to a model whose window fits the load (or drop the pin to inherit the session model), after verifying which agents carry the pin and their spec sizes
confidence: medium (mechanism inferred from 3/3 failures + 1/1 fix; exact window limits unverified)