fix: Remove evaluation metric key from schema which failed on some LLMs

[jsonbailey]

fix: Improve metric token collection for Judge evaluations when using LangChain
fix: Include raw response when performing Judge evaluations

---

Note

Medium Risk
Updates the judge structured-output contract and parsing logic, plus changes LangChain structured invocation to return parsed/raw data and token usage; this could affect downstream integrations expecting the old schema or metrics behavior.

Overview
Judge evaluations now use a fixed structured-output shape. EvaluationSchemaBuilder no longer bakes evaluation_metric_key into the schema; Judge now expects top-level {score, reasoning} and keys the parsed result by the config’s metric key, failing the evaluation when the response doesn’t parse into a valid score/reasoning.

LangChain structured invocations now capture more telemetry and handle Bedrock better. invoke_structured_model uses include_raw=True, returns only the parsed payload, surfaces raw_response, extracts token usage from either usage_metadata or response_metadata, and treats parsing_error as a failed structured call; provider mapping now routes bedrock and bedrock:* to bedrock_converse and injects Bedrock’s foundation provider parameter when needed.

^{Written by Cursor Bugbot for commit a303c3d. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}