One builder, even across clusters

The first production run of the dedicated build process surfaced a coordination gap: it and a starting server each believed they had won the build claim, and each ran the full build. Nothing corrupted — every build output is content-addressed, so both wrote the same artifacts — but the server paid the build's cost that the dedicated process exists to spare it.

The claim decision is now a deterministic election. Registrations get a short convergence window so that candidates from different clusters see each other before anyone decides; then every decider computes the same winner from the same data, so even concurrent decisions agree. A dedicated build process always outranks a serving server — which is what makes servers stand down exactly when a dedicated builder is present, while remaining their own fallback when none is. Takeovers of a departed builder stay immediate, and per-part claims stay instant.