fix(providers): mark image-capable OpenRouter models as vision-capable

[mousamax]

Fixes #481.

What was broken: all four curated OpenRouter entries use _AGENTIC, which is vision=False. OpenRouter declares image input for Kimi K2.6 and Llama 4 Maverick. Since engine.py:1085-1096 swaps image parts for a placeholder on non-vision models, the image never left the machine and the model replied that it could not see it.

Fix: both entries now carry vision. They use an inline ModelCapabilities rather than _AGENTIC_VISION, which would also set pdf=True - PDFs stay unverified over the compat surface, so these fall back through pdf_support.py like together:moonshotai/Kimi-K3. GLM-5.2 and DeepSeek V4 Pro are text-only upstream and are unchanged.

Source: GET https://openrouter.ai/api/v1/models, architecture.input_modalities, checked 2026-08-09.

Verification:

New test pins all four ids against the declared modalities.

$ pytest tests/test_providers.py -q
26 passed in 0.60s
$ pytest tests -q
1172 passed, 1 skipped in 34.16s

Live, with an OpenRouter key on Kimi K2.6 via OpenRouter. Before:

After: