fix: score a released-seat wake once, not twice

[tonicmuroq]

Found while running an e2e of the retain/release hibernate+wake cycle on a staging cluster. Metric-only defect — no state is corrupted, which is presumably why it went unnoticed.

What happens

The wake-completion arm (reconcileWake, case waking && vmLive(main)) returns ctrl.Result{Requeue: true}, and controller-runtime serves the next pass from the informer cache — which can still carry the status this pass just overwrote. The arm runs a second time. Its own hint-clear is visible by then, so hint == "" and the same landing gets scored placement=pool.

Measured, not inferred

Two full release cycles on a staging CocoonSet (operator v0.3.2, vk v0.3.7). Both wakes landed on the hinted node:

counter	after cycle 1	after cycle 2	per cycle
slot_release_pods_deleted_total	1	2	+1 ✅
slot_release_wake_total{placement="hint-node"}	1	2	+1 ✅
slot_release_wake_total{placement="pool"}	1	2	+1 ❌
restored on … log lines	2	2	2 per wake

2/2 reproduction. A 100%-hint-node result reports as a 50/50 split, which makes the placement ratio — the entire point of the metric — unusable. pool is only ever labelled when the hint is empty or mismatched; the landing node was correct both times, so the empty hint is the only way in, confirming the re-entry.

The fix

The hint is the wake's own in-flight marker: reconcileSuspendRelease sets it before any pod delete (that patch blocks the delete loop, so a released wake always has one) and this arm clears it exactly at completion. Gate the scoring, the log line and the clear on it still being present — only the pass that owns the completion counts.

DeleteManifest deliberately stays outside the guard rather than moving under it: if a hint ever went missing, a guarded version would leave the :hibernate tag behind, and vk's derived restore-from-hibernate path treats a lingering tag as "a wake was lost here". Keeping it unconditional costs one 404-tolerant registry round trip on the re-entry and keeps that safety net.

Test

TestWakeScoresPlacementOnceAcrossStaleStatusReentry drives the completion arm twice, the second time with the hint cleared and the status still Waking — exactly what the Requeue hits live. Verified it actually catches the defect: with the guard removed it fails with pool delta=1, the same symptom measured in production.

go build ./..., go test ./..., make lint (linux + darwin) all green.

Also

Drops Every release wake restores via registry pull from the metric's help text. Since vk-cocoon v0.3.7 keeps the node-local snapshot across a seat release, a hint-node landing restores locally — snapshot_pull_duration_seconds_count stayed 0 across both cycles above.