fix(CC-572): pr-gate synthesis retry removes (not truncates) the result path

[screenleon]

Summary

Diagnosed from repeated production protocol failures during the CC-571 gate saga: apply_patch failures ("invalid patch: multiple operations target <file>", "invalid hunk", "Failed to find expected lines") kept recurring on the PM-synthesis correction retry, in both sequential and parallel gate modes — 4 consecutive protocol failures across that saga's rounds.

Root cause: the reviewer-protocol retry (which never showed this failure) always writes to a brand-new path (reviewer-<name>-<ts>-retry1.md) — always an unambiguous "Add File" for whatever patch tool the executor uses. Both synthesis retries instead re-dispatch to the SAME fixed $OUTPUT_FILE path. The sequential retry already truncated it (: > "$OUTPUT_FILE") before this fix, and still hit the identical failure — confirmed by direct trace inspection that a path which still EXISTS on disk, even at 0 bytes, can lead the executor's patch tool to choose an "Update File" operation (which locates existing content to edit) instead of "Add File". An update against empty/rewritten content then fails outright, since there's no content to locate — a hard patch-tool failure unrelated to the actual synthesis content. The parallel retry didn't even truncate; it left the first attempt's full content in place.

Fix: remove the path entirely (not truncate it) before both retries, via a new shared gate_clear_retry_target helper next to gate_dispatch_command. This forces an unambiguous "Add File" the same way the already-reliable reviewer-retry path's brand-new filename does. The helper extraction (found via the mandatory reuse/simplify pass — all three review angles independently flagged the same duplication) also collapsed two ~15-line near-duplicate root-cause comments into one canonical explanation.

Verification

• Added a test-only capture hook (CODEX_GATE_CAPTURE_OUTPUT_EXISTS_DIR) that records, at the very start of each synthesis dispatch, whether the result path exists on disk at that instant — and two regression tests (one per mode, sharing one parameterized helper) asserting the first dispatch sees "exists" (reviewer content already written) and the retry sees "absent".
• pmctl gate run --executor codex --policy generic — Final: GO, first attempt, all four reviewers approved (didn't happen to trigger a synthesis retry itself, so this specific gate run didn't exercise the fix live — but the mechanism is verified against the actual historical failure traces plus white-box regression tests).
• All existing synthesis-protocol (14) and sequential-protocol (4) tests still pass.
• tests/bin/run-all-tests.sh — 104 passed, 0 failed, 0 skipped (full sign-off; caught and fixed a lint-test-docstrings violation on the two new test wrappers along the way).

🤖 Generated with Claude Code

https://claude.ai/code/session_0177ds9nxEMtsh3bwpBxNHYx