feat: v0.9 chat slim checkpoint (81M, CTIBench 28.9%)

v0.9 chat-tuned checkpoint, slimmed from 973MB to 324MB by dropping
optimizer + grad scaler state. Inference-only.

Architecture: 6 layers, d_model 768, 12 heads, RoPE + SwiGLU + RMSNorm.
Pretrain corpus: 273M tokens (PRIMUS + CWE + OWASP + RFCs + fact-QA).
Chat-tuned with the chat-v3 SFT recipe.

Bench (chat-tuned, debiased text-scoring):
CTIBench MCQ full (n=2500, 2-perm): 28.9%
in-repo CTF MCQ eval (n=30): 59.2%
SecQA (n=210, external): 39.3%
free-form fact recall (n=50): 1/50 (at floor for the
81M rung; the next
move is parameter
count, not corpus)

Used by the Ghostgim/ghostlm Space at runtime via hf_hub_download.

best_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72030d38702bf4738364b2d40cb4f780d7b68c17f4a46ad4c4a0f55908449f06
+size 324416181