Phase A serving speedup: zero-copy state-cache expand (no deepcopy per decision)

Serving-path speedup: state-KV LRU cache + suffix diet (see inference/bench_serve.py)

typical-medium (tm1b): weights, evals, inference package, manifest