Build job can be simplified.

[Robert Bruce Park]

The new status system refreshes all silo statuses every 30 minutes. This is effectively like running a WATCH_ONLY build on every silo every 30 minutes.

Because that's now happening, WATCH_ONLY builds are effectively obsolete. We could just do away with BuildBase.watch_phase entirely, such that the build job just merges merges, prepares packages, generates diffs, then exits, allowing the migration script to report the build status after the fact.

This would also make diff generation more reliable as they aren't currently generated if there's a build failure. Then we could drop diff regeneration from the start of the publish job.