perf(ds4): overlap q5 verification and stabilize 16K

[davide221]

Summary

• batch q=5 hyper-connection pre/post/output work and drafter-feature capture across all verifier lanes
• place profiled experts by the predicted two-owner critical path instead of aggregate hit rate
• use chunked heterogeneous prefill scratch, proactive graph-cache eviction, and a sticky 1K long-context shape on no-VMM devices
• add reproducible overlap tracing and exact-context qualification controls

Exact-context qualification

Configuration: q=5, four active routed experts, 14,350 MiB hot-expert budget, main/peer rate 4.4, two warm-ups plus three measured 128-token replies per context.

Context	Client decode median
2K initial	75.818 tok/s
4K	74.530 tok/s
8K	69.898 tok/s
16K	62.685 tok/s
2K after 16K	76.703 tok/s

All 25 requests completed with the expected deterministic response hash. Peak main-GPU VRAM was 31.089 GiB. The post-16K 2K measurements were 76.685-76.727 tok/s.

A same-budget placement A/B moved 116 experts to the peer and measured 76.636 tok/s at rate 3.8, within 0.1 tok/s of the qualified rate. A post-batching trace found 16-22% simultaneous kernel-busy time in steady decode windows; target verification remains 63.7 ms of a 74.4 ms speculative step.

Validation

• test_deepseek4_unit: pass
• test_server_unit: 322 passed, 0 failed
• full 2K -> 4K -> 8K -> 16K -> 2K model-backed sweep: pass
• shell syntax and Python compile checks: pass
• git diff --check: pass