load latest weights + enable trained MTP draft head (use_writer_mtp)

Cortex-A 0.5: latest phase-2 weights, bundled tokenizer + hf_transfer, backbone-active peak inference, enhanced UI, removed train/analysis info

v32: full-cond writer + MTP head structure, top-p+rep-penalty sampling, fix diffusion-era prose

v30: AR writer (diffusion removed)

v29: z_scale=1 fix

v29: z_scale=1 fix

v27: 8L block-residual decoder, bf16 rope fix, 4-step default

v27: 8L block-residual decoder, bf16 rope fix, 4-step default

v25: arithmetic chunk ids

v24: block-causal decoder + per-chunk t (leak closed); clean-past denoise

fix jit tracer: max_chunks from cache shape

0.6 latent-diffusion: cortex/latent_diffusion.py

0.6 latent-diffusion: cortex/model.py

0.6 latent-diffusion: cortex/losses.py

0.6 latent-diffusion: cortex/config.py

Cortex-A 0.5 live chat: KV-cache decode (CPU) + streaming + tok/s + progress panel