Target-only Qwen3.8 W4 sampled decode selects mixed-dtype packed GDN

Native Windows CUDA target-only decode for the Qwen3.8-27B W4A16 gate model fails before producing a sampled token when temperature is 0.7 and top-p is 0.8. The production path reaches vt::GdnPackedDecode with mixed_qkv/a/b/out not sharing one FP16/BF16/F32 dtype and throws at src/vt/ops.cpp:2496. The same target model runs through speculative DFlash2 verification because that step takes a different GDN recurrence route. Reproduction evidence: J:\vllmcpp-dflash2-w4-runs\73-target-only-repro-t07-p08-1x4. Owning row: SPEC-REJECTION. The fix must preserve the packed-decode dtype predictor/producer invariant and add a target-only production regression test.\n\nFOLLOWING_AGENTS_PROTOCOL\n\nFollowing-Agents-Protocol: true\nAI-Assisted: true\nAssisted-by: AGENT:GPT-5 [CODEX]

Start at the mixed-dtype failure in src/vt/ops.cpp:2496 and use the reproduction evidence at J:\vllmcpp-dflash2-w4-runs\73-target-only-repro-t07-p08-1x4 to trace the target-only production path into vt::GdnPackedDecode. Compare the packed-decode dtype predictor and producer, then add a target-only production regression test covering the Qwen3.8-27B W4A16 sampled decode case. Done means the path preserves the shared dtype invariant and produces a sampled token without the rejection.