Investigate KONI short-input pre-EOS drift across RNGD and GPU #261

Problem

KONI-8B live RNGD prefill→GPU decode differs from native GPU references before EOS for128-token outputs. Transport/lifecycle tests pass, so exact-output performance claims need a separate numerical boundary investigation.

Evidence

Checkpoint0172bf7e73977ae213613529f3d2d5400a33f678, BF16, RNGDwidth8192, GPU4paged lanes/context9216/page64/cacheoff. Completed24batches: all32-output conditions match;128-output differs at index64(short128input) or98(long8192input), same at c1/4, no EOS128001/128009 in either128-output sequence.
For8192input, nativeGPU chunk8192 and retainedRNGDKV both reproduce live128tokens; chunk256 reproduces originalnative. For128input, native128 and HF128prefill→PDBatcher agree with originalnative. HF8192prefill padded by repeating lasttoken→trim128KV→PDBatcher matches neither originalnative nor liveRNGD, diverging at index11. Thus physical shape affects arithmetic but padding alone does not explain shortRNGD output.
Raw local201-koni-matrix-a.json, local201-koni-parity-audit.json, local201-koni-diagnostic.json, koni-short-padding-diagnostic.json; experiment2026-10-24-live-rngd-gpu-preflight.md. Allfour GPU lifecycle/decode-isolation evidence is in #260; it does not establish cross-hardware numerical equivalence.

Scope

Capture live short-prompt RNGD KV, verify replay matches live output, compare identical forced-prefix logits under BF16/FP32 and layer KV differences. Preserve failing original comparisons; no fixture-specific token forcing in production.

Done when

Identify and validate an actual correction across held-out prompts, or document a numerical compatibility boundary with broader quantitative evidence. Single-fixture parity or lifecycle success alone is insufficient.

Owner

Codex; maintainer review. Related #201, #219, #221, #248; analogous Qwen4 #256 and Gemma #233.