v1.6 — OpenRouter, Claude & GPT-5.5

Added

• OpenRouter provider support: models can declare provider: openrouter
  in models.yml and are routed through an OpenAI-compatible OpenRouter
  client. New openrouter_api_key / openrouter_api_base config options.
• Anthropic Claude models via OpenRouter: Claude Opus 4.8, Claude Sonnet
  and Claude Haiku.
• OpenAI gpt-5.5 (via OpenRouter).

Changed

• Chat models are now dispatched by their type in models.yml instead of
  hardcoded model-name lists, so adding a model is a config-only change.
• Token counting falls back to the o200k_base encoding for models unknown
  to tiktoken (e.g. Claude).
• The /settings model picker lays buttons out in rows of two to stay within
  Telegram's per-row inline-button limit.
• Image understanding is now driven by a vision: true flag in models.yml,
  so any vision-capable model (GPT-4o, GPT-4o mini, GPT-5.5, Claude) can read
  images — no longer limited to GPT-4o / GPT-4 Vision.