Builds can run in their own disposable process

The same portal binary — and therefore the same image and the same framework fingerprint — can now run as a dedicated build process instead of a serving portal. In bake mode it joins no serving cluster (it gets its own cluster identity, so build work and live traffic cannot contend or starve each other), runs the coordinated build against the shared stores, publishes the GO signal, and exits.

Exiting is the point: everything a full build allocates — measured at roughly two gigabytes for a complete sweep — is reclaimed the moment the process ends, instead of living on inside a serving pod. And because a build's outcome must never be ambiguous, the process's exit code is its verdict: success, a faulted sweep that verified nothing, or a misconfiguration that ran no build at all.

Serving pods keep their own build capability as a fallback — the claim on the build root arbitrates cleanly between a dedicated build process and a pod baking for itself, so running the build ahead of a rollout is an acceleration, never a requirement.