orchestrate: a mid-turn reviewer kill loses the whole review, and no gate detects the scratch files it leaves undeclared in the worktree #385

Found during orchestrate run prd352-20260803-015333 (PRD #352, 12 slices, all passed).

Symptom

The reviewer-deep on slice #355 was terminated by a usage limit (idleReason: "failed", failureReason: "You've hit your session limit"). Two distinct losses followed.

1. The review itself was lost entirely

The reviewer had clearly done substantial adversarial work — it had authored a 208-line probe suite exercising recoverSliceProgress and resolveSliceProgressPath — but produced no envelope, so none of its conclusions survived. The envelope is emitted as a final flourish, which is precisely the moment a turn-budget or usage cutoff destroys.

2. It leaked a scratch file into the worktree

test/zz-review-probe.test.ts was left behind, self-labelled:

// TEMPORARY reviewer probe — deleted before the slice is handed back

The reviewer intended to delete it and died first. Because the file is untracked and undeclared, finalize_slice would not stage it (it stages exactly the declared union) — but it would be collected by vitest and change the orchestrator's pre-merge test count, and it survives into any later inspection of the worktree as apparent slice content.

Impact

Loss 1 is expensive but recoverable — the run-state subState: verified checkpoint is exactly the right resume anchor, so a re-spawned reviewer redoes only the review, not the investigation or implementation. That part of the design worked.

Loss 2 is the silent one: nothing in the pipeline detects an undeclared file appearing in the worktree between verify_changeset (step 4a, before the reviewer) and the pre-merge gate (step 5a, after it).

Workarounds used

Removed the probe file manually before re-running the gate; re-spawned the reviewer with two added instructions — write the envelope early rather than last, and delete every scratch file.

Suggested fixes

• Make envelope emission incremental or checkpointed rather than terminal. The same argument ADR-0017 makes for the slice progress record applies to the envelope itself. At minimum, the bundled subagent definitions should instruct every role to write a provisional envelope as soon as it has a defensible verdict and revise it, rather than composing it once at the end.
• Re-run verify_changeset (or an equivalent undeclared-file check) AFTER the reviewer returns, not only after the implementer. Today the reviewer is trusted to declare its own edits with no worktree cross-check — an asymmetry with the implementer that has no stated rationale.
• Have the reviewer's prompt template mandate scratch-file cleanup. Currently nothing forbids leaving scratch files behind.

Relationship to existing issues (read before deduping)

• Partially overlaps orchestrate: bound the bundled investigator-standard (haiku) exploration so it emits its envelope before truncating #346 on the envelope-first ask, but orchestrate: bound the bundled investigator-standard (haiku) exploration so it emits its envelope before truncating #346 is scoped narrowly to investigator-standard (haiku) over-exploring to truncation, and its primary fix is an exploration budget for that one role. This issue asks for envelope-first across all roles, on the different trigger of an externally-imposed usage cutoff that no exploration budget can prevent.
• Not orchestrate: distinguish a transient (API/socket) subagent crash from a genuine missing-envelope FAILED slice #334 / orchestrate: recover a missing-envelope implementer whose worktree holds near-complete work (extend continue-in-place beyond incomplete) #267. Both are about recovering a missing envelope (retry, or continue-in-place). The undeclared-scratch-file half here is not a recovery question at all: it is a missing verification step between two existing gates, and it fires whether or not the reviewer dies.
• Not orchestrate: implementer/reviewer subagents assert a non-existent "advisor pass" in result envelopes (false provenance, ADR-0009) #285 (false advisor provenance in envelope prose) — a content-honesty issue, not an emission-timing or worktree-hygiene one.

Environment

orchestrate 1.6.0 (plugins/orchestrate/.claude-plugin/plugin.json:3 on development) plus the PRD #352 umbrella branch (slice-executor delegation layer). Run prd352-20260803-015333; slice #355, reviewer-deep.