workflow-fix: gotchas rc=134 — in-process streaming-datasets release

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #952 (emitting agent: experiment-implementer).

Goal

Extend the #654 rc=134 gotcha entry in .claude/rules/gotchas.md with the in-process prevention: deterministically release streaming datasets objects (del ds; gc.collect()) after consumption, never os._exit(0).

Workflow gap

• Bug observed: A datasets streaming IterableDataset surviving to interpreter shutdown SIGABRTs the process (rc=134, "terminate called without an active exception") AFTER all work completed and the final sentinel was written — deterministic in the pinned env; a clean GCE pod run would be classified as a workload crash by the EXIT trap (#952 r2, bisected with a bare 15-row streaming loop).
• Why it is a workflow gap: The existing #654 gotcha covers only the WRAPPER side (artifact-check before treating rc as fatal); no entry prescribes the in-process prevention, so every new streaming-load_dataset call site re-hits the false-failure class.
• Confidence (emitter): high

Proposed change (candidate diff sketch — refine in planning)

- **HF `datasets` ... exit `rc=134` (SIGABRT) ... AFTER the work already completed ...**
+ ... RULE (wrapper side, unchanged): artifact-check before treating rc as fatal.
+ IN-PROCESS PREVENTION (#952 r2): a streaming IterableDataset surviving to
+ interpreter shutdown reliably triggers this abort — release it deterministically
+ after the consuming loop (`del row, ds; gc.collect()`; verified rc 134 -> 0).
+ Never mask with `os._exit(0)`. Worked example: run_952.phase0_verify.

Scope / surfaces

• Primary target: .claude/rules/gotchas.md
• Grep the workflow surface for the pattern before editing (grep -rln 'rc=134\|SIGABRT' .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.
• 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/gotchas.md
• fingerprint: 35a510cc017e

target_file: .claude/rules/gotchas.md bug_observed: A datasets streaming IterableDataset surviving to interpreter shutdown SIGABRTs the process (rc=134, "terminate called without an active exception") AFTER all work completed and the final sentinel was written — deterministic in the pinned env; a clean GCE pod run would be classified as a workload crash by the EXIT trap (#952 r2, bisected with a bare 15-row streaming loop). why_workflow_gap: The existing #654 gotcha covers only the WRAPPER side (artifact-check before treating rc as fatal); no entry prescribes the in-process prevention, so every new streaming-load_dataset call site re-hits the false-failure class. proposed_change: Extend the #654 rc=134 gotcha entry with the in-process fix — deterministically release streaming datasets (del ds; gc.collect()) after consumption, never os._exit(0). confidence: high related_task: #952