Qwen3.8-27B on two Blackhole P150A

An optimisation campaign for single-stream coding inference of `Qwen/Qwen3.8-27B` on two P150A cards linked by QSFP-DD (TP2), using DSpark speculative drafting, fused MLP and custom Tensix kernels, with CI workflows that replay each experiment on hardware and on `ttsim`. The author reports 118 committed tok/s at 4K context and 50 tok/s at 64K, and states plainly that the 200 tok/s target has not been reached. These are offline runtime tests, not a serving benchmark, and reproducing them depends on a Docker image that is not published to a registry. The docs include a Blackhole tuning playbook, a gotchas list and a record of fixes contributed upstream to tt-metal.