With one launch per gate and the state copied to the device and back, a
whole job measured about as fast, or slower on the GPU, between 18 and 22
qubits on the fleet's GPUs (22 qubits on the laptop GPU: 159 ms on the CPU,
202 ms placed). A planner must not route by default to a path that is not
better, so `auto` now places on the CPU on every host and looks for no
member: a job that does not ask runs as it always did, its result unnamed.
A member runs a job only when the job asks: `accelerator` takes the first
certified member that holds the state, and one that fails during the run
hands the job to the next, the receipt listing each that failed
(`fell_back_from`, now a list, each with how it failed); when every one
fails the job is refused (J013). A member named by id or key is alone.
Discovery, certification, the receipt's fabric block, the split and
`wai-quantum fabric` are as they were; `fabric_members` states `auto`
instead of a threshold. A test with stand-in devices holds the fall-back to
the CPU's bits and to the schema.
The hardware legs ask for `accelerator` and check that auto names nothing;
both replayed on jetson-hub and framework-128gb. Docs, help, both MCP
servers' descriptions, the guide and the changelog say so; agent docs
re-blessed.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
780b0526ce