Rollouts no longer wait on a pod that never started

When a server is deleted while it is still starting up, the cluster keeps a record of it that no health check ever revisits — nothing is watching a member that never finished joining. Startup compilation reads that record to decide whether the server currently responsible for the build is still working, and it read "still starting" as "still running". So if the responsible server disappeared mid-boot, every other server waited for a build that would never happen: a rollout could sit for 25 minutes or more, with no error anywhere to explain it.

This is fixed. A member that never finished joining is now treated as "no information" rather than "running", which hands the decision back to the liveness stamp the build already writes — so the work is picked up by the next server as soon as that stamp goes stale, and the rollout completes. A server that is genuinely mid-start and working keeps its build, exactly as before.

The same change also starts the liveness stamp earlier, at the beginning of the build rather than after its planning phase, so a long build can no longer look abandoned while it is running.