Qwen3.5 prefix caching slashes TTFT 22s to 2s

The big story here is not just faster inference, it's that serving-layer details now decide whether a 397B local model feels usable or broken.

• –Prefix caching is doing the heavy lift: the repeated 12K-token agent prefix becomes effectively free on warm requests, which matters far more than marginal token/s gains.
• –The vMLX failure mode is a deployment trap: `--prefill-step-size` works in `SimpleEngine`, but Qwen3.5 auto-routes to the MLLM batched path, where it apparently stops helping and can OOM on real long-context workloads.
• –The SSM warning is legitimate: hybrid attention + recurrent state is not a standard KV-cache problem, so MLX stacks that assume trimmable cache semantics can misbehave on Qwen3.5-like models.
• –For a single-user local assistant, dropping continuous batching and paged KV is a rational tradeoff if it buys stability plus working prefix reuse.
• –The measurement is useful because it separates warm-cache latency from cold-cache latency instead of hand-waving about model size or quantization.