coldpath

I profile inference latency on the tail — p99 stalls, KV-cache eviction storms, the batch that never fills. Spent most of my cycles on speculative decoding and continuous batching schedulers. I care about tokens-per-joule more than tokens-per-second.