[gspark-etri]

Background

KONI-8B live RNGD→GPU completes requests but differs from nativeGPU beforeEOS on128-token outputs. Lifecycle success does not explain numerical drift. This PR retains failed comparisons and controlled probes without changing production sampling.

Changes

Add original live matrix/audit, matched-prefill and padded-prefill controls, shortKV repeat/replay metadata, identical-prefix BF16/FP32/head-onlyFP32 probes, and unforced output-head precision controls. Include portable forced-prefix entry point and exact local research scripts. No production code changes.

Verification

Runtime32ce5d5, checkpoint0172bf7e73977ae213613529f3d2d5400a33f678; A10080GB+RNGD, SDK2026.4, GPU Torch2.13. Original inputs128/8192,outputs32/128,c1/4,3repeats,4paged lanes/context9216/cacheoff, RNGDwidth8192.

• Original32-output matches;128-output differs atindex64(short) or98(long), beforeEOS.
• Long nativeGPU full8192prefill and retainedRNGDKV both reproduce live128tokens; chunk256 reproduces originalreference.
• Short GPU128prefill/import matches native; padding to8192 matches neither native nor live. Padding-only explanation fails.
• Short liveKV exported twice: firsttoken andBF16K/V bit-identical; replay reproduces original live128tokens. Native original128tokens also reproduced.
• Common64-token prefix: BF16native logits3940/8712 tie14.875/14.875; retainedKV14.8125/14.9375. FP32native/import both select8712. Output-head-onlyFP32 also resolves that inspected tie.
• Unforced128-token controls reject head-onlyFP32 as a generalfix: KONIshort firstdifference14 (original64), Qwen4long13 (matchedBF16original107); KONIlong matchedprefill remains equal. No fixture-specific production tie-breaking added.
  Hardware command/results are in experiments/2026-10-24-koni-numerics261.md. All raw complete flags/counts verified; portable CLI syntax/help and scoped sensitive/diff scans passed. CLI adaptation itself was not hardware rerun; equivalent original script was.

Compatibility and limits

Diagnostic evidence, not a universal numerical/quality fix. Firstdifference measures equality, not semantic accuracy. FP32native changes prefill/decode arithmetic; importedKV remainsBF16.17MBshort and1.1GBlong KV dumps remain local; capture script/dependencies documented. Local research scripts need their environment paths adapted; matching Qwen control inputs are in #257. No equivalent-output speedup claim.

Related Issue

Refs #261,#201,#256. Does not close numerical investigation. Stacked on #259; draft for review and broader validation. Maintainer merges.