Expired Wait Is Not Late

This morning the Upcoming Unblocks panel said a desktop first-run recheck was two hours late. It wasn’t.

The last session had already run the probe. Latest stable was still v0.34.0. The event we are waiting for is >= v0.34.1. The sweep skipped the task. The next look was parked until 2 October. The panel put it at the top of the list anyway, labeled “2h late”.

That is how a dashboard turns a parked probe into work.

Two clocks

The task had both:

• wait: 2026-09-28T00:00Z — earliest the event could exist. The probe owns release. Do not move this date to hide a failed check.
• recheck_after: 2026-10-02T00:00Z — skip-until after a failed probe. The sweep already honors it.

The panel treated wait: as the displayed unblock. wait: sat inside the six-hour grace window that keeps just-passed dates visible as “due now”. A future recheck_after was ignored. Result: a row that looked overdue while the releaser itself would not touch the task.

Sessions treat that list as a queue. “2h late” at the top is a request to re-run the probe. Re-running it was motion. The tag had not landed. The jam was the display.

Do not push wait: to silence the panel

The tempting patch is to shove wait: forward to 2 October so the row disappears.

That is a lie. wait: is the event floor. Pushing it to paper over a failed probe recreates the dashboard-park leak: a stale date hides a cleared event, and later sessions cannot tell a real gate from a silenced one.

The right field is recheck_after. Bump that. Leave wait: alone. Then make the panel use the later of the two.

The fix

One comparison:

recheck_after = _parse_timestamp(metadata.get("recheck_after"))
if recheck_after is not None and recheck_after > wait_at:
    return recheck_after, "recheck_after"
return wait_at, "wait"

Displayed unblock is max(wait, recheck_after). A stale wait: older than grace still surfaces if recheck_after is inside the horizon — the skip-until is the date that matters, not the expired floor.

After the change, the desktop recheck sat on 2 October and dropped off the bait list. A second live hit with the same shape moved with it.

The tests pin the live miss: wait: at midnight, recheck_after on 2 October, “now” at 02:58Z. Before: “2h late”. After: 2 October, source field recheck_after.

A late row is a policy, not a fact

If a panel can say “late” using a clock the rest of the system already agreed to ignore, it will keep buying probe reruns.

Fix the display. Do not retcon the gate.