providers=[...] includes image and music generation models as candidates

[TheCoder30ec4]

Problem

fetch_catalog() keeps every model whose output_modalities contains "text". Image and music generators also list text as an output, so they become routing candidates for text tasks.

Checked against the live catalog: 7 of the 24 models picked for providers=["google"] are generators:

google/gemini-3.1-flash-lite-image    google/gemini-3.1-flash-image
google/gemini-3-pro-image             google/gemini-3.1-flash-image-preview
google/gemini-3-pro-image-preview     google/lyria-3-pro-preview   (text + audio, listed at $0)
google/lyria-3-clip-preview           (text + audio, listed at $0)

The lyria music models show as $0, so Jev sees them as the cheapest option for simple text tasks.

Where

src/model_router/catalog.py, the filter in fetch_catalog().

Suggested fix

For routing text tasks, keep only models whose output_modalities is exactly ["text"]. Explicit models=[...] could still allow anything, if we want an escape hatch.

Done when

• Router(providers=["google"], ...) excludes image and music models
• Test in tests/test_catalog.py with a text+image and a text+audio model
• README's "Choosing candidate models" mentions the rule

[Ansh0305]

I can solve this