Refresh default model catalog for 2026 lineup

[bouclem]

What

Refreshes the default model catalog for the 2026 lineup. Every name in
defaultModelsOfProvider now has a real backing capability entry, so
nothing silently falls through to the 4K-context, no-tools default.

Models added

OpenAI (openAIModelOptions)

• gpt-5.5, gpt-5.5-pro
• gpt-5.4, gpt-5.4-pro, gpt-5.4-mini, gpt-5.4-nano
• effort slider extended to include xhigh for the gpt-5.x line
• pro variants are reasoning-only (no none option)
• fallback recognizes dated IDs like gpt-5.4-mini-2026-03-17

Anthropic (anthropicModelOptions)

• claude-opus-4-7 (1M ctx, adaptive thinking only — reasoningCapabilities: false)
• claude-sonnet-4-6 (1M ctx, extended thinking)
• claude-haiku-4-5-20251001 (200K ctx, extended thinking)
• fallback ordered most-specific last so claude-opus-4-7 wins over
  the more general claude-opus-4 rule

Gemini (geminiModelOptions)

• gemini-3.5-flash, gemini-3.1-pro-preview, gemini-3-flash-preview
• replaces the previous always-null fallback so these now resolve

xAI (xAIModelOptions)

• grok-4.3-latest (1M ctx, configurable reasoning)

DeepSeek (deepseekModelOptions)

• deepseek-v4-flash and deepseek-v4-pro (1M ctx, 384K max output)
• v4-pro cost uses list price, not the current 75% promo
• replaces the previous always-null fallback

Groq (groqModelOptions)

• openai/gpt-oss-120b (131K ctx, openai-style tools)

Defaults updated

• New names listed in defaultModelsOfProvider for openAI, anthropic,
  xAI, gemini, deepseek, groq
• Old entries kept in *ModelOptions for OpenRouter / OpenAI-compat /
  legacy users — only the dropdown defaults changed
• Dropped grok-4-1-fast and groq/compound from defaults (not
  suitable for IDE coding workflows)

Verification

• TypeScript clean, no diagnostics on the modified file
• Walked through every name in defaultModelsOfProvider to confirm it
  resolves to a real entry (exact match or fallback) instead of the
  4K default

Notes

• DeepSeek v4-pro pricing is the published list price. The current
  75% promo isn't encoded since promos expire.
• Cache pricing for cache_read on Anthropic 4.x not included — the
  docs page didn't surface those numbers in the model table.