Fix/geak revalidation dispatch

[lishuoshuo-amd]

Problem

A GEAK e2e win only reaches the headline after a same-harness rebench revalidates it. Across 68 audited sessions, 19 of 26 eligible candidates (73%) never completed that rebench, and the quality gate never once fired — the candidates were not rejected on merit, the rebench never ran to completion.

Two defects caused it:

1. The phase boundary cancelled the rebench. It runs as an explore task and no phase after KERNEL allows that kind, so KERNEL_AGENT -> SWEEP cancelled the task KERNEL had just dispatched — 475ms after creation in the sampled session.
2. The idempotency key was a fixed string. "geak-revalidate" is unique per session, so from macro-cycle 2 onward create_or_return_existing returned the settled first-cycle row, which KERNEL read as rebench_unavailable. All 20 affected sessions sat at macro_cycle >= 2.

Making the rebench survive and retry then exposes four more:

3. A rebench spared at every boundary also survives into CLOSE, where it runs a full serve-and-bench that starves the post-opt roofline of the GPU lane and can rewrite the headline after the report is written. The wall-clock timeout path is worse: it never reaches the CLOSE sequencer, so it cancels queued work but leaves a running rebench writing back through the grace window and the slot claiming a rebench is still coming.
4. Per-cycle keys plus no in-flight check allow two live rebenches; the orphan's result is taken at face value.
5. A cancelled attempt cannot be replaced within its cycle, because its key returns the cancelled row.
6. A candidate abandoned without revalidation produces no fact and no warning in the report — the session reads as if nothing was found. This predates the branch: rebench_unavailable was already invisible.

Fix

1. cancel_queued_not_allowed takes a spare_queued predicate; the GEAK policy spares the rebench at every boundary except CLOSE. Deny-list, not allow-list: the phase set changes and a missing allow-list entry silently restores the bug, while a surplus costs one wasted bench that the gate in (4) refuses anyway.
2. Keys are scoped per macro-cycle.
3. Both wind-down paths — the CLOSE sequencer and the wall-clock closing phase — share one drain that cancels queued and running rebench, then settles the slot. Settling is driven by state, not by what the caller just cancelled, since the phase boundary has already cancelled the queued case by then; it edits the verdict fields in place so the candidate's self-reported numbers survive for the report.
4. _enqueue_geak_revalidation returns early while a rebench is in flight, and a result is applied only when geak_pending still tracks the finishing task. Applies to all four 2b decisions, including the inconclusive branch, which otherwise reached the 2a harness replay and cleared the tracked slot. Ignored results emit a geak_rebench_result_ignored observation.
5. Keys step past settled attempts (-r1, -r2, bounded per cycle). Pruning the explore family settles the slot instead of leaving KERNEL held open.
6. rebench_unavailable and the new rebench_cancelled both emit a warning naming the drop reason, in the final report and the mission view.

GEAK policy lives in orchestrator/phases/geak_rebench.py. TaskRegistry stays generic, receiving only the spare_queued hook and a find_by_idempotency_key lookup.

RCA: Desgin-DOCS/hyperloom-geak-revalidation-dispatch-rca.md

Tests

26 regression tests in test_geak_revalidation_dispatch.py covering per-cycle keys, re-dispatch after cancel, the SWEEP spare and the cancel on both wind-down paths, duplicate-dispatch suppression, prune settling, audit fields surviving a settle, orphan and late rejection across all four decisions, and report rendering for dropped candidates.

pytest src/hyperloom -n 12
# 14339 passed, 21 failed

The 21 failures are pre-existing and environment-induced (credentials / gateway auth), each reproduced on a clean origin/main worktree. test_objective.py::test_run_closing_phase_writes_report_on_time_exhausted is a pre-existing flake driven by its 6ms max_minutes; measured 8 runs at pass=6 fail=2 on both this branch and origin/main.

Breaking changes

None. rebench_cancelled is a new additive geak_pending.status; readers treat unknown statuses as non-pending. Sessions on the old key resume correctly — the slot holds a task id, which matches directly.

[github-actions]

CI E2E report — ✅ Succeeded

item	value
result	✅ Succeeded
model	Qwen/Qwen3-0.6B (dense)
resources	1× GPU, TP=1
PR branch	fix/geak-revalidation-dispatch
commit	a9f93fa6e1e8bf051551ee0476f8426a43764f37
session_id	68c45c93-83ea-455b-9578-1053e2847d77
queue → dispatch	0s
run time	167m 13s
total	167m 13s

details

[ZhengGong-amd]

Approved