fix(models): thinking config for gateway model IDs + new model generations

[philipph-askui]

Why

uses_adaptive_thinking() classifies models by exact claude- prefix match. That breaks in two ways:

1. Gateway model IDs — Bedrock (anthropic.claude-opus-4-8, us.anthropic.claude-...-v1:0), LiteLLM (anthropic/claude-...), and Vertex (claude-...@date) IDs don't match, fall into the budget_tokens path, and get a 400 on every ≥4.7-generation model (budget_tokens is removed there).
2. Future models — the adaptive list is an allowlist, so every new Claude release 400s until the list is updated.

Additionally, AndroidAgent hardcodes thinking={"type": "disabled"} + temperature=0.0, both rejected on newer generations (sampling params removed from Opus 4.7 / Sonnet 5 / Fable 5; explicit disabled rejected on Fable 5).

What

• Normalize before matching: model IDs are reduced to their claude-... core, so gateway wrappers classify like bare IDs.
• Inverted classification: adaptive is the default; only the frozen legacy budget_tokens families (2.x/3.x, Haiku 4.5, Sonnet 4/4.1/4.5, Opus 4/4.1/4.5) keep the integer budget. Unknown future models work by default.
• New helpers accepts_sampling_params() (False from Opus 4.7 / Sonnet 5 / Fable 5 onward) and supports_disabled_thinking() (False on Fable 5 / Mythos 5), plus make_non_thinking_settings() used by AndroidAgent — same behavior as before on older models, valid requests on newer ones.
• EffortLevel gains "xhigh" (supported from the Opus 4.7 generation onward; recommended for agentic use).

Notes

The C# SDK carries a port of this module; the mirroring fix is askui/csharp-sdk#16 — the two should land in step to keep parity.

Test plan

• pytest tests/unit/models/test_thinking.py — 54 passed, including new matrices for Bedrock/LiteLLM/Vertex ID shapes, sampling/disabled-thinking classification, make_non_thinking_settings() per generation, and xhigh
• ruff check / ruff format --check clean on changed files
• askui.android_agent imports cleanly

🤖 Generated with Claude Code