Scope GPU layer and CUDA buffer evidence to the latest target model load, before a separate speculative drafter starts loading. Require that target to report full offload so a smaller draft buffer cannot reject a valid Bolt installation or hide partial target offload.

Cover the Bolt target/drafter log sequence, zero and partial target offload, and a stale successful load followed by a partial load.