Only allow generative (text/vision) models in the chat model picker

[Lazarus-931]

Problem

The chat model picker is a denylist (LocalModel.isEligibleForLanguageModelPicker = !speechToText && !textToSpeech), so non-generative models a user downloads — encoders like google-bert/bert-base-uncased / FacebookAI/xlm-roberta-base, embeddings, ASR, time-series — still appear and can be selected as the chat model. Selecting one and starting the server pre-loads it as text_generation; mlx-vlm/mlx-lm has no matching architecture, so startup aborts:

ERROR - Error loading model google-bert/bert-base-uncased: Model type bert not supported
ModuleNotFoundError: No module named 'mlx_lm.models.bert'
Application startup failed. Exiting.  (exit status 3)

The app is then effectively bricked — the server won't come up and the menu bar is unresponsive, with no in-UI way to recover.

Fix

• Picker becomes an allowlist of generative models: text and multimodal vision-language models both qualify (.text or .vision); only models with no generative capability — encoders, embeddings, ASR, forecasting, pure image-gen — drop out. Multimodal/VLM models are not excluded.
• More robust generative detection: infer .text from causal/generation architectures (*ForCausalLM / *ForConditionalGeneration / *LMHeadModel) in addition to the existing name heuristics, so legitimate chat models (Phi, SmolLM, …) aren't accidentally filtered out.
• A stale non-generative selection no longer bricks startup: when the configured chat model's local config.json shows only non-generative architectures, its --model pre-load flag is dropped (load-on-demand) so the server still starts, with a log note. The check is conservative — any uncertainty leaves the model untouched.

Screenshots to follow.

[Blaizzy]

Please double check this doesn't affect images which are supported in the chat version in images tab

[Blaizzy]

Also, it doesn't affect omni models like Gemma 4 which might use that tag

[Lazarus-931]

yeah @Blaizzy, great point about omni models, added a small patch for that, for the img page, i think we can wait to merge this until a new pr for folding the img feature into the chat page. i think @lucasnewman said it'll be a follow up work

[lucasnewman]

yeah @Blaizzy, great point about omni models, added a small patch for that, for the img page, i think we can wait to merge this until a new pr for folding the img feature into the chat page. i think @lucasnewman said it'll be a follow up work

I added that in #93 so you should be good -- model availability is type-aware now for non-language models (image, stt, tts).

[lucasnewman]

LGTM but see comment about making sure to pick up all possible model roots.

[Lazarus-931]

yeah @Blaizzy, great point about omni models, added a small patch for that, for the img page, i think we can wait to merge this until a new pr for folding the img feature into the chat page. i think @lucasnewman said it'll be a follow up work

I added that in #93 so you should be good -- model availability is type-aware now for non-language models (image, stt, tts).

wait, confused, the release shows sep img and chat page