Module serving_state 

The serving control plane as one sans-IO state machine over the algebra: every transition performs exactly one program evaluation, or placement copies only, and the per-layer cache lives inside the variant a transition returns rather than in a cache object patched from outside.

The tests below prove the control flow (drafting, acceptance counting, cache placement, rollback-to-resume) against a fake target function instead of a model program.

Enums

ServingFsmError

ServingState

Which evaluation shape is legal right now. An enum whose transitions consume self states that directly; no existing proxima primitive (pipe, cache or config) expresses it more cheaply.