[antonreshetov]

Adds OpenRouter as a dedicated cloud provider for the AI assistant, as discussed in #964. A generic OpenAI-compatible URL stays a separate follow-up.

• Uses @openrouter/ai-sdk-provider@^2.10 (compatible with ai@6) through the existing SDK path: streaming, tool calling and reasoning replay. The endpoint is pinned to https://openrouter.ai/api/v1 and the requests use the existing guarded fetch.
• The model list only includes tool-capable models (supported_parameters=tools) and hides :batch variants, which return 404 in chat.
• /models on OpenRouter is public, so the key is verified with GET /key before listing models.
• Reasoning responses without signatures (DeepSeek, Qwen, etc.) no longer fail with upstream: undefined-valued provider metadata is normalized before replay validation.
• A rejected key after saving the connection now shows a dedicated message (all providers). HTTP 402 (no credits) maps to the quota error.
• sdkModel is now an exhaustive switch, so a new SDK provider without a model fails typecheck.
• Docs: OpenRouter row in the AI provider table and the home page mention.