feat(qwen36): add all-C request-granular adaptive verification

[Graffioh]

Summary

This draft adds request-granular adaptive speculation for every supported active concurrency (C <= 16). It does not classify prompts as code/chat and it has no fixed concurrency cutoff: each request may use the configured speculator or remain autoregressive, based on measured useful-token goodput.

The implementation is split into a reusable planning policy and a concrete DDTree adapter:

• Common, model-neutral request, confidence-scout, verifier-work, and work-menu types.
• One adaptive profile per exact (speculator, shape, proposal nodes, verifier rows) work key, with a shared exact-C AR baseline.
• Separate hard executor capacity and adaptive efficiency limits, so forced requests remain correct without making ordinary adaptive routing too wide.
• Confidence and accepted-token yield stay isolated per work shape; scout time is profiled separately from verifier-route time.
• The Qwen concurrent engine supplies DDTree work menus and executes the selected DDTree subset together with AR peers in the existing packed mixed route.

The common selector accepts one configured speculator per call. DDTree, DSpark, and future speculation methods can implement the same adapter contract without changing the ranking policy. Simultaneously racing multiple speculators would be a separate portfolio policy and is intentionally not claimed here.

Decision process

At each decode step:

1. The active speculator adapter enumerates the exact verifier work shapes executable for each request.
2. Cold requests receive a bounded confidence scout. DDTree derives conditional prefix survival from draft probabilities; another speculator may provide the same signal from a confidence head.
3. The selector orders requests by expected useful output, then progressively replaces that prior with target-verified accepted-token yield.
4. It compares absolute predicted useful-token goodput for measured route widths against the naturally observed exact-C AR cost.
5. Only the profitable request subset speculates; every other request stays on AR.

There is no prompt keyword heuristic, fixed code/chat label, or per-model offline performance table. Confidence expires by generated-token progress, and costs continue adapting through bounded EWMAs. New/refilled cohorts, draining tails, transient scout failures, AR-peer protection, and Always/Never request semantics have explicit safeguards.

DDTree integration

• DDTree confidence scouting returns a reusable proposal artifact tied to request ID, slot, seed token, committed position, and top-k shape.
• A selected artifact is reused only when all identity/position checks still match.
• Only scout work actually consumed by the selected route is charged to that route; failed or unselected scouts cannot poison verifier profiles.
• DDTree scouts are currently serial because draft KV is request-local. The batch scout interface permits a future DSpark confidence-head adapter to fuse that operation.
• C <= 3 retains the established full DDTree verifier shape while routing requests independently.
• C = 4 profiles the established full shape before a compact fallback.
• C > 4 offers bounded compact shapes first, then wider work only after the cheaper exact profile loses.
• Closed cohorts keep the AR-peer guard. Refill can relax it only when the exact work has broad executable coverage or stable target-verified evidence.
• Forced speculation may use every eligible request up to hard executor capacity; it is not capped by the adaptive efficiency limit.

This PR wires DDTree end to end. DSpark is not executed concurrently by this PR; adding it later means implementing its scout/menu/executor adapter against the common contract.

Strix Halo results

All GPU workloads were pinned with HIP_VISIBLE_DEVICES=1; runtime logs confirmed Radeon 8060S / gfx1151. The R9700 was not used.

Protocol: Qwen3.6-27B Q4_K_M, local DFlash Q4_K_M draft, greedy decoding, fresh server per workload and mode, 24-token warmup, then two measured runs with exactly 64 output tokens per request. All requests completed with the exact token count and zero failures.

Mixed matrix

C	workload	adaptive tok/s	AR tok/s	tok/s delta	mean-latency delta	measured steady route
2	1 code + 1 chat	22.184	16.833	+31.79%	-46.12%	DDTree used
4	2 code + 2 chat	31.029	31.459	-1.37%	+1.39%	AR
5	2 code + 3 chat	37.623	37.742	-0.32%	+0.32%	AR
8	4 code + 4 chat	53.410	53.645	-0.44%	+0.44%	AR
16	8 code + 8 chat	77.970	78.232	-0.34%	+0.34%	AR

At C=2, both requests used DDTree in the measured runs: the code request accepted 63/63 speculative tokens per run, while chat accepted 22 and 19. This is a real aggregate win for that fixed mixed workload, but it is not evidence that the controller selected only the code request.

For C >= 4 in this matrix, every measured request reported zero DDTree steps and exactly C * 63 target forwards. The controller therefore converged safely to pure AR; these rows demonstrate bounded overhead/parity, not a high-concurrency DDTree speedup. Sub-1% differences should be treated as noise-level parity, while the C=4 -1.37% is the clearest residual scout cost.

Targeted cohorts

C	workload	adaptive tok/s	AR tok/s	tok/s delta	observation
3	2 code + 1 chat	24.683	18.783	+31.42%	steady DDTree use
4	4 code	30.746	31.523	-2.46%	one measured cold probe, then AR
5	5 code	37.574	37.715	-0.37%	AR
6	3 code + 1 structured + 2 chat	42.129	42.133	-0.01%	AR / exact parity
6	6 code	42.049	42.119	-0.17%	AR / parity

A separate forced C=5 smoke test produced 5 x 24 tokens with zero failures, and all five requests recorded DDTree steps and accepted tokens. This confirms that hard capacity is not accidentally limited to the adaptive lane count.

Current conclusion: the controller preserves the real C=2-3 DDTree gains and avoids a steady regression at higher concurrency, but this Strix/Qwen/DDTree configuration does not yet demonstrate profitable steady speculative work at C >= 4. A cheaper/fused DSpark scout or a faster verifier shape can reuse the policy and be evaluated independently.

Validation

• Release dflash_server and test_speculation_goodput build: passed.
• Adaptive policy suite: 484 checks passed.
• Focused CTest selection (server_unit, recurrent snapshots, engine contracts, batch plans, speculation policy): 375/375 passed.
• Forced C=5 DDTree capacity smoke: passed on Strix Halo.
• Final P0/P1 code review: no findings.
• git diff --check: passed.

Stack

This remains the existing draft PR #617 and keeps its current stack/dependency relationship with #605. The latest cohesive implementation commit is cd6cf301.