Prefix reuse saves prompt work and shared storage when state matches

Compatible requests can share cached prompt state, saving both prefill work and KV memory. Calculate those savings separately, then check whether routing to a warm cache is worth the extra queueing.

14 MIN · PLUS