experiment: validate KONI long-prefix RNGD handoff to standard decoders #219

Problem

KONI8B has long standalone GPU/ATOM coverage but RNGD-to-ATOM evidence only uses an older short external-cache graph. The selected standard B1/context9216/k1 artifact has not consumed long RNGD KV.

Hypothesis

The existing Llama3.1 FullPrefillEngine can produce KONI2K KV with its existing RoPE scaling and geometry. Export group tensors on the host into the same portable format as the Qwen matrix, then compare GPU and standard ATOM continuation.

Scope

Validate the exact KONI checkpoint/geometry; reuse FullPrefillEngine kernels. Add a small host-export adapter and a repeatable long-input runner, exercise2K first. Permit the ATOM checker to select its already-validated k1 artifact. Record numerical disagreement and compilation failures. No default engine or shared SDK changes, and no overlap/performance claim from these correctness runs.

Done when

Actual-device repeatability and decoder handoff have evidence, or a reproducible failure identifies the missing work. GPU/ATOM output and KV readback comparisons are reported separately. Full four-model goal remains open.

Owner

Codex, maintainer review. Related #201, #208, #217 and #194.