The two rungs

Address provenance (D149): an address that resolves is now counted as evidence about the GPU-address assumption rather than discarded as a precondition met, and counted apart from whether the shader translated. The failure path had to be split to do it - "nothing is mapped there" and "the shader is not translatable" are evidence about different things.

The submission entry point (D200): submit_at takes an address and a length, which is the shape a real call site has.

643 workspace tests green.

Surprises

• Rung one as stated could not be built. "Submissions contribute to a run's progress block" assumes submissions; nothing anywhere calls Pipeline::submit, so a progress line would have said the same thing on every run forever. What could be built is the seam: the entry point a shim will call, tested, with the address measured on the way in. The shim itself sits on the loader side - thread 1 already identified the import that marks a submission - so this is the half that can exist before the two meet.
• The measurement was already being taken and thrown away. Every submission tested the address assumption and reported only pass/fail of the shader behind it. Nothing new had to be computed; it had to be kept.