update: Eagle3 finetune ep9 weights (KL loss, post-norm data, 29K samples, 10 epochs)

Finetuned from nvidia/Kimi-K2.5-Thinking-Eagle3 weights on K2.5 API log dataset.
Offline eval: pos0=88.2%, pos1=74.0%, pos2=61.9% (val_postnorm_eval, 466 samples).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f00f89f509445ae7a602399025672d0fb388f1de24038b83811e34a1b7fcc24e
+oid sha256:eff87daf70cd1d3bde2b83b5b422096f6a05c29de6811db197f0d6a52da6f959
 size 3682399920