CD failed on main: run 36237075862

[github-actions]

CD failed on main: run 36237075862 for commit 5497c81a1f22ad19b7df7f560c062123dbe011e6 (entered via push).

Triage the failed job in the linked run - which of the two things CD delivers is broken depends on WHICH job went red:

• an IMAGE job / promote / verify-images: no complete image set was published. Publication is all-or-nothing, so every self-updating install stays on the PREVIOUS image until a CD run promotes the full set.
• preflight / notify-platform-update: the images may be fine, but the release EVENT did not reach the control instance, so no node repo re-bakes against this commit until its own schedule poll (Release broadcast has never delivered: notify-platform-update succeeds, Hosting/PlatformBuilds stays empty, no satellite dispatch since 08-22 #2235). Read the job's ::error:: line - it names what to provision or what refused the delivery.
• satellite-compat (a matrix leg) / satellite-compat-image: the set IS published and the release event WAS sent - what failed is the per-build measurement that a satellite's NodeTypes still compile against it (the counterweight to the daily rebuild, 2026-09-12). A red leg means THIS BUILD BROKE that satellite: its types compiled against the set the fleet is on and do not against this one (a satellite that was already broken, or an unestablished baseline, is an advisory on the run page, never a red). The lines under Compatibility below name the types. No image and no publication is missing.

The hourly reconciler will re-attempt main's HEAD on its own (bounded at 3 attempts per commit, logged here); you do NOT need to force a merge to trigger delivery. This issue CLOSES ITSELF when a CD run reports a successful heal (#3176) - a later failure files a fresh issue rather than appending here, so this one's title and age stay true to its own cause.

[github-actions]

✅ Healed cdda28d (run) — main's HEAD now has the complete image set (portal 3.0.0-ci.9406), verified against ACR. Installs pick it up on the build-completion event, or within one safety-net check interval if that event never reaches them (#2494).

Closing this issue, because the delivery hole it records is filled. A later failure files a FRESH ci-failure issue rather than appending here (#3176), so an alert's title names its own cause and its age is the outage's age. Reopening is never needed — and closing also hands the reconciler a fresh heal budget.