fix(llm): recognize claude-opus-4-8 as a reasoning model on Bedrock

[kimnamu]

HUMAN:

I reviewed this change locally against the pinned litellm 1.84.1, confirmed the Bedrock cross-region reasoning gap, and verified the regression tests fail before / pass after the one-line fix.

• A human has tested these changes.

AGENT:

---

Why

LiteLLM 1.84.1 recognizes anthropic/claude-opus-4-8 but not its Bedrock cross-region ids (bedrock/{us,eu,apac,global}.anthropic.claude-opus-4-8-v1:0), so on Bedrock the model falls through to the non-reasoning branch and leaks temperature/top_p → Anthropic 400. This mirrors the fable-5 gap fixed in #3642; opus-4-7's Bedrock ids are already covered by LiteLLM metadata.

It deliberately uses REASONING_EFFORT_MODELS (strips temp/top_p only), not EXTENDED_THINKING_MODELS — adding it there is what #3427 did and #3441 reverted, because that branch injects the legacy thinking.type=enabled block + interleaved-thinking header that Anthropic now rejects for opus-4-8.

Summary

• Add "claude-opus-4-8" to REASONING_EFFORT_MODELS (model_features.py) — one data line, mirrors the fable-5 fix (fix(llm): recognize claude-fable-5 as a reasoning model #3642).
• Membership is case-insensitive substring, so the single entry covers all Bedrock CRIS prefixes (us./eu./apac./global.).

Issue Number

Fixes #3694

How to Test

Run the added offline unit tests (no network):

uv run pytest tests/sdk/llm/test_model_features.py -k opus_4_8
uv run pytest tests/sdk/llm/options/test_chat_options.py -k bedrock_opus_4_8

Before/after evidence: with the source line reverted (git stash push -- openhands-sdk/openhands/sdk/llm/utils/model_features.py), the new tests FAIL — temperature leaks and _supports_reasoning_effort("bedrock/us.anthropic.claude-opus-4-8-v1:0") returns False (5 failed, 36 passed). Restoring the line → 174 passed. Local pre-commit (ruff + pyright) passes.

item	Before	After
_supports_reasoning_effort("bedrock/us.anthropic.claude-opus-4-8-v1:0")	❌ False	✅ True
temp/top_p in request for that id	❌ sent (→ Anthropic 400)	✅ stripped
anthropic/claude-opus-4-8 (already worked)	✅ True	✅ unchanged
legacy thinking.type=enabled block / beta header	n/a	✅ NOT injected (asserted in test)
public API / signatures	—	✅ unchanged (data line only)

Type

• Bug fix
• Feature
• Refactor
• Breaking change
• Docs / chore

Notes

This contribution was prepared with the help of an AI agent (Claude Code); a human reviewed the change, rationale, and test results before submission.