perf(qwen35): improve concurrent AR prefill latency

[Graffioh]

This PR is based directly on main. Its complete GitHub diff contains only server runtime code, server documentation, and server-side tests. The benchmark tooling and detailed measurement methodology are isolated in #596.

What this changes

• sizes the Qwen3.6 target-step graph for mixed packed-prefill plus compact-decode schedules, fixing the C16 abort caused by a ninth recurrent segment;
• permits bounded prefill-only traversals while preserving paused decoder token, KV, and recurrent state exactly;
• preserves FIFO packed-cohort behavior by default, with wide-cohort round-robin available as an explicit opt-in for future C32+ work;
• exposes bounded prefill-burst and idle-token budgets with fail-closed parsing;
• removes the unconditional 15-byte SSE tool-parser delay from ordinary requests with no declared tools, while preserving split reasoning-marker detection and conservative tool/stop handling.

Runtime behavior

The screened default remains DFLASH_PREFILL_FIRST_BURST_STEPS=0. Prefill-first bursts can improve a native prefill tail but delay decode, so they remain opt-in. DFLASH_PREFILL_COHORT_ROUND_ROBIN is also opt-in; the default remains FIFO plus the existing intra-cohort partial-budget cursor.

Dynamic graph sizing retains the old 16,384-node capacity through eight recurrent segments, then adds 2,048 nodes per additional segment. The observed C16 mixed schedule requires 16,741 nodes, so it receives an 18,432-node graph.

Strix Halo validation

The benchmark tooling lives in #596. One-repeat screens ran exclusively on gfx1151 / Radeon 8060S.

• C16 short: Luce goodput 30.96 versus llama.cpp 22.67 tok/s; max TTFT 23.404 versus 29.561 s.
• C16 long: Luce goodput 6.08 versus llama.cpp 5.19 tok/s; max TTFT 158.331 versus 180.800 s.
• Native Luce C16 prefill throughput was broadly flat across context cohorts (about 349-363 tok/s); latency growth primarily followed total prompt work.

The SSE follow-up used the exact same long-C16 selected prompt set (54,148 prompt tokens, 16/16 valid, exactly 64 output tokens/request). Median TTFT moved from 140.464 s to 123.210 s (-12.3%); max TTFT was essentially unchanged at 157.854 s. This is a one-repeat screen, so the direct emitter regression test and the reduced visible-minus-native delay are the stronger causal evidence.

Validation

• scheduler plan: 139 checks passed;
• sequence-engine contract: 13 checks passed;
• graph-sizing regression passed;
• complete server unit suite: 362/362 passed;
• final HIP dflash_server built and linked for gfx1151;
• production long-C16 smoke completed 16/16 without graph or scheduler faults;
• rewritten server tree is byte-identical to the previously tested and benchmarked server tree;
• complete PR diff against main contains only 13 server/ files.

Review notes

• omitted decoder rows do not append or commit KV and do not advance recurrent state;
• a bounded prefill burst always yields to the complete live decoder cohort after N pauses;
• idle epochs reset burst and cohort state;
• ordinary no-tool content streams immediately, while partial <think> / </think> fragments remain boundary-safe.

[Graffioh]

Emitter latency follow-up (commit abae692)

Upstream behavior checked before implementation:

• vLLM buffers only request stop-string suffixes in its incremental detokenizer; reasoning/tool parsing is handled separately.
• SGLang emits clean incremental text immediately and mainly delays incomplete UTF-8.

Luce had imposed its 15-byte tool-syntax holdback even for ordinary requests with no declared tools. This follow-up makes ordinary content immediate while retaining split reasoning-marker detection; requests with tools or custom stops retain the existing conservative buffering.

Validation:

• server unit suite: 362/362 passed, including immediate short content and a split opener regression
• production gfx1151 HIP server linked
• exact long-C16 Strix screen: same selected prompt hash, 54,148 prompt tokens, 16/16 valid, 64 output tokens/request
• median TTFT: 140.464 s -> 123.210 s (-17.254 s, -12.3%)
• max TTFT: 158.331 s -> 157.854 s (essentially flat)
• median visible-minus-native-prefill delay: 28.339 s -> 22.367 s

This is a one-repeat screen. Prefill order varied between runs, so the direct unit regression and visible-minus-native gap are the cleaner causal evidence. Six C16 requests became visible about 0.33 s after native prefill; ten still waited about one 22.37 s traversal because their first sampled token produced no visible text. That remaining delay is a separate scheduler/detokenizer follow-up.