fix: clamp Bedrock thinking effort to the LiteLLM-accepted ceiling (#737)

[ElegantLin]

Stacked on #736 (base = test/599-bedrock-effort-parity) — it reuses the _wire_output_config_effort end-to-end helper introduced there. Will rebase onto main once #736 merges.

Summary

BenchFlow accepts BENCHFLOW_BEDROCK_THINKING_EFFORT values xhigh/max, but LiteLLM 1.88.0rc1's Bedrock Converse transform rejects them for adaptive Claude 4.8+ with a BadRequestError (issue #737). So the #598 "MAX thinking" config errored at request time instead of running at the highest supported effort.

Fix

Clamp a requested effort to the LiteLLM-accepted ceiling (high) in both injection paths, matching the existing garbage→high default so a MAX-thinking run proceeds at the real maximum rather than crashing:

• Route config (litellm_config._bedrock_thinking_effort) — bakes the clamped value into config.yaml.
• Standalone proxy patch (litellm_bedrock_patch) — clamps the os.environ override too. It is deployed into the sandbox alone (cannot import benchflow), so it carries its own in-sync copy of the ladder; a comment in each points at the other.

Test plan

• Route param clamp: xhigh/max → high; updated the two pre-existing tests that encoded the bug (max threaded through unchanged) to assert the clamp.
• Proxy-env path: a run-level max override clamps in the real wire payload without raising.
• Drift guard: every requestable effort, after the clamp, is accepted by the REAL litellm Converse transform — if a future litellm changes its accepted set, this fails loudly instead of shipping a rejected value mid-run.
• Both clamps verified fail-then-pass when individually neutered.
• 87 passed across the bedrock/litellm slice; unrelated agent-CLI reasoning_effort tests unaffected; ruff + format + ty clean.

Closes #737

🤖 Generated with Claude Code