feat(qwen35): gate speculation by measured acceptance

[Graffioh]

Summary

• Stack on feat(qwen36): complete concurrent serving implementation with speculation #605 and carry forward its packed mixed DDTree/AR verification, fused replay, and direct-commit execution substrate.
• Use one measurement-derived score per request: optimistic during probation or stale re-probing, otherwise accepted-token EMA.
• Sort requests by score and admit the speculative prefix only while each additional lane improves predicted goodput, the route remains at least 5% above AR, and AR-peer latency stays within 110%.
• Keep only three persistent request measurements: EMA yield, observed rounds, and progress at the last successful speculative round. There are no content labels, feature vectors, trained classifiers, modes, or demotion counters.
• Preserve per-request adaptive, ar, and speculation controls plus the fixed-policy and forced-oracle test overrides.

Verification

• HIP Release configure for gfx1151
• dflash_server and affected engine/unit targets build successfully
• focused gate, mixed-path, scheduler-contract, snapshot, and decode-mode suite: 9/9 passed
• complete server unit suite: 366/366 passed
• whitespace and superseded-policy audits are clean

Notes

This is stacked on the current #605 head (e3cee32). The M3 model-backed throughput/TTFT benchmark matrix has not been run yet and remains follow-up validation for this draft.