Refresh the model catalog, with checksums and context sizes

Do. Replace the 2024 models with current open coding models that fit consumer GPUs, chosen by our own tool-calling test and license check (candidates: Qwen 3.6/3.8 and Qwen3-Coder-Next, Devstral Small 2, GLM-4.7-Flash, gpt-oss-20b). Add sha256 (the downloader already verifies it), context length and a recommended quant (dynamic Q4_K_XL/Q5_K_XL; no sub-4-bit for tool calling).

Spec §5.2 #7.