fix(anthropic): use token thinking for effort levels on models without adaptive support

[Sayt-0]

Summary

Switching the active model to Haiku 4.5 (or any model without adaptive-thinking support) and setting a reasoning level via Shift+Tab made every subsequent request fail with HTTP 400. The Shift+Tab cycle stores the level as an effort-based thinking_budget, and the Anthropic provider routed any effort budget into the adaptive-thinking API (thinking.type=adaptive with output_config.effort) regardless of model.

Issue expectations

Expectation (#3362)	Handled by	Status
Do not set adaptive thinking on models that do not support it	resolveThinkingBudget converts effort/adaptive budgets to token budgets on unsupported models	yes
Use the regular thinking options instead	thinking.type=enabled with budget_tokens derived from the effort level (effort.BedrockTokens)	yes
Reasoning should work properly	verified end to end (see Testing)	yes

Change

• pkg/modelinfo: add SupportsAdaptiveThinking (Opus 4.6/4.7/4.8, Sonnet 4.6, the Claude 5 families, Fable, Mythos). It is a superset of RejectsTokenThinking, so a model that rejects token budgets always reports adaptive support.
• pkg/model/provider/anthropic: replace coerceAdaptiveThinking with resolveThinkingBudget, which resolves both directions: token to adaptive for Opus 4.6+ (unchanged), and effort/adaptive to a token budget for models without adaptive support (new). adjustMaxTokensForThinking now operates on the resolved budget so a converted token budget gets the correct max_tokens headroom.

Testing

Unit tests (added/updated) cover the effort-to-token fallback, the adaptive path on supported models, the previously untested plain effort level, and the SupportsAdaptiveThinking superset invariant. lint, vet, and gofmt are clean.

Verified against a live gateway (provider: anthropic, proxied to Bedrock):

Case	Path taken	Before	After
Haiku 4.5 + effort high	token 16384	HTTP 400	passes, reasoning works
Haiku 4.5 + effort medium / low	token 8192 / 2048	not applicable	passes
Sonnet 4.6 + effort high	adaptive (no regression)	not applicable	passes

Notes

The Bedrock provider (provider: amazon-bedrock) already routes plain effort levels to token thinking for non-Opus-4.6+ models and is not affected by this issue. It retains a narrower edge case (an explicit thinking_budget: adaptive on a non-adaptive model), which is out of scope here and can be addressed as a follow-up. If a different resolution seems more appropriate, it can always be changed.