daily-fix: phase idempotency + inter-phase schema assert

Overview / Motivation

Auto-filed by the /daily 2026-07-25 problem sweep from task #1689. This is the money item of the day: a non-idempotent phased dispatcher re-ran a paid Anthropic-API phase on every crash-fix relaunch, roughly four times.

Goal

Make phase-level skip-if-output-exists mandatory for multi-phase dispatchers (a BLOCKER-grade code-review gate when a phase makes paid API calls), require the CONSUMER phase to assert its inter-phase JSONL contract before any model initialization, and record the interpreter-finalization early-exit gotcha.

Workflow gap

Three failures on #1689 (session 5c5a89e8), all at phase boundaries:

1. No idempotency ⇒ repeated paid API spend. issue1689_dispatch.sh all has no skip-if-complete guard. Phase A (corpus 3800 rows + render 79,800 rows) re-ran on every one of ~4 relaunches, and haiku_u2 (3 conditions × 3800 rows of real Sonnet calls, smoke=False) re-ran at least twice — the second time purely because a downstream vLLM phase crashed. Crash-persisted Phase-A artifacts already existed on HF under issue1689_partial/. CLAUDE.md's checkpoint-per-phase rule exists; nothing enforced it here.
2. Unasserted producer/consumer schema ⇒ a wasted GPU cycle. After the renderer produced 79,800 rows across 21 conditions, onpolicy died with ValueError: The decoder prompt cannot be empty at vllm/v1/engine/processor.py:488. Root cause (found in R8): chat-framing rows emit messages: [...] but no prompt_text, and the consumer called row.get("prompt_text") — 33% of rows resolved to empty. The failure landed after vLLM initialization, so it cost a pod cycle rather than seconds of CPU.
3. Interpreter-finalization crash turned a completed phase into a pipeline abort. [corpus] done: scanned=5132 kept=3800 — output written — then Fatal Python error: PyGILState_Release … Python runtime state: finalizing, a C-extension atexit race across 203 loaded modules (torch/scipy/sklearn/pandas/ pyarrow). Under the dispatcher's set -euo pipefail, the nonzero rc from a phase whose WORK was complete killed phases B–E. R6 fixed it by adding an explicit process exit to 8 entrypoints.

All three share one shape: a phase boundary with no contract. (1) has no "already done" contract, (2) has no input contract, (3) has no "work complete ≠ exit code" contract.

• Confidence (emitter): high on all three (each carries the task's own epm:failure / implementer-report text). Medium on where the enforcement belongs.
• verified-at-filing: grep -rn --exclude-dir=worktrees 'checkpoint-per-phase' .claude/rules/code-style.md CLAUDE.md → the rule EXISTS (CLAUDE.md § Code Style names "checkpoint-per-phase" in the code-style pointer list), which is the point: the rule is present and was not enforced, so the gap is a missing GATE, not missing guidance — clause (c) context read confirms the existing text is advisory prose, not a reviewer-binding item. Incident text quoted above read from #1689's own epm:failure markers (reason: vllm-empty-prompt, reason: python-interpreter-shutdown-crash) and the R5/R8 implementer reports, not from recall. Landed-fix history check git log --oneline --since='7 days ago' -- .claude/agents/code-reviewer.md .claude/rules/code-style.md → no idempotency or inter-phase-contract gate landed. (2026-07-25)

Proposed change (refine in planning)

  code-reviewer gate (kind: experiment diffs adding/altering a phased dispatcher):
+ (1) every phase skips when its declared output artifact already exists
+     (or an explicit --force is passed). BLOCKER when the phase makes paid
+     API calls or holds a GPU; CONCERN otherwise.
+ (2) the CONSUMER phase asserts its inter-phase contract at the top —
+     required fields non-empty, row count, fail-loud drop report — BEFORE
+     any model/vLLM/accelerator initialization.

  gotchas.md entry:
+ a phase entrypoint importing torch/scipy/sklearn/pyarrow should exit the
+ process explicitly after flushing; a nondeterministic interpreter-finalization
+ crash otherwise converts a COMPLETED phase into a whole-pipeline abort under
+ `set -euo pipefail`.  (#1689: Phase A wrote its output, then died in finalize.)

Scope / surfaces

• Primary target: .claude/agents/code-reviewer.md (the two gate items).
• .claude/rules/gotchas.md (the finalization entry) and/or .claude/rules/code-style.md § checkpoint-per-phase — the planner picks one home and cross-references, rather than duplicating.
• Consider whether the idempotency gate belongs partly in planner.md §9 (declare each phase's output artifact so the reviewer has something to check against). A gate that cannot name the expected artifact will degrade to a judgement call.
• Do NOT hard-code #1689 specifics; the schema mismatch is task-specific, the unasserted contract crossing a phase boundary is the recurring class.

Constraints / invariants

• The idempotency requirement must not break deliberate re-runs: --force (or an equivalent) stays first-class, and a partial/corrupt output must not read as "done" (prefer a completion sentinel over bare file existence — the project's own .claude/rules/ Monitoring rule already bans keying "done" on file existence).
• The consumer-side assertion must fail LOUD (no silent drop, no default fill) — CLAUDE.md § Critical Rules "Fail fast".
• scripts/workflow_lint.py --check-references / --check-asks pass; ruff passes; agent-spec size ratchet stays under budget.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

• workflow_fix_target: .claude/agents/code-reviewer.md
• fingerprint: 4cb38a3bc0db
• Source: /daily 2026-07-25 transcript sweep, session 5c5a89e8 (#1689) @ 2026-07-26T05:29Z/06:20Z (re-runs), 05:40:29Z (vllm-empty-prompt), 02:31:13Z (finalization crash).