fix(speculative): Qwen3.5 MTP verification drift breaks temperature-0 parity #77

When speculative decoding verifies a Qwen 3.5 MTP draft block, the target model's per-position logits can disagree with single-token decode, which flips the accept/reject decision and breaks the temperature-0 byte-equality invariant (greedy speculative output should be identical to drafter-less greedy decode).

Root cause

The Qwen 3.5 MTP target-verify pass verifies the whole draft block in one forward. Its full-attention layers run the batched fused causal attention over the multi-token block, whose reduction order differs from single-query decode attention. The resulting per-position logit drift can flip the target's argmax at a verify position, diverging the accept/reject decision from the drafter-less greedy pass. Qwen 3.5 is especially sensitive because its gated attention and quantized projections amplify the drift through the gate.

Fix

Thread a target-verify flag through the Qwen 3.5 verify pass. When verifying a multi-token block, compute attention per query position: query i attends only to keys and values up to prefix_len + i + 1 with no mask, exactly as during single-token decode. The decode and prefill paths are untouched: the flag defaults to off and all existing call sites pass off, so non-speculative inference is byte-identical and pays zero overhead.

Acceptance criteria

• The Qwen 3.5 MTP target-verify pass produces per-position argmax tokens identical to single-token decode at the same absolute positions.
• Non-speculative decode and prefill remain byte-identical (no behavior change when the verify flag is off).
• An executable regression test pins the verify-versus-decode argmax parity.