Coverage and gaps

Request or gate	Evidence and outcome
Exact migrated configuration	Identity; matching model/image, explicit NCCL and host-stack differences
Historical speed comparison	Four matched prompt cells, C1 n3 each; descriptive variable-output baseline
Controlled output and unique cache	Strict scout 0/3; unique natural 1/10 canary compliant; dependent cells stopped
Direct protocol gates	Disabled 9/10; enabled passes
Coding quality	Five cases x3, 15/15
Image/OCR corpus	Six cases x2, 12/12
Extended context	Native sweep 128/150; 9 empty length-terminated and 13 incorrect visible answers; non-monotonic curve
Agentic recovery and long sessions	Deep profile, 30/30
Matched repository-agent smoke	Official SWE grader, resolved 1/1; 22 requests versus historical 11
Sustained reliability	Endurance, 60/60
Routed protocol gates	Disabled 9/10; nominal 380K HTTP413; 260K passes; enabled reasoning evidence fails projection
Post-run state	Verified unchanged serving state; routed smoke/JSON passed

The Windows arm is retained evidence rather than a newly rerun installation. Strict controlled-output performance qualification did not pass. Cache state, driver and transport prevent an OS-only causal claim. C1 is the deployed engine ceiling; higher concurrency, full SWE-bench, video and speculation retuning are outside this deployed-profile comparison. No profile promotion is authorized.