experiment: validate Qwen3 RNGD prefill and ATOM handoff #197

Problem

The current research RNGD prefill accepts only Qwen2. Qwen3-0.6B ATOM standalone matches CPU on the three tested prompts, but no heterogeneous handoff is validated.

Evidence

Qwen3-0.6B has hidden_size1024,16query heads,8KV heads,head_dim128,28layers, Q/K normalization and no QKV biases. The current Qwen2 graph assumes query width equals hidden size and biased projections.

Scope

Add a separate, strictly scoped research Qwen3 prefill runner preserving Q/K normalization and independent query width; verify KV/logits against CPU, then host-mediated ATOM admission using a dedicated compatible paged-KV compiled artifact. No transport ABI changes or automatic model-family enablement.

Done when

Recorded actual-device numerical and handoff checks, with remaining disagreements explicit. Unsupported quantized/sliding/MoE configs rejected before device work.

Owner

Codex (experimental implementation; maintainer review).
Related: #194, #195.