feat(evaluation): expose evaluationReferenceInputs on EvaluatorInput

[jariy17]

Issue

Closes #539

Description of changes

The Lambda event for code-based evaluators carries fields that EvaluatorInput dropped, so evaluator functions had no typed access to them. This adds the missing ones (see input schema).

Ground truth (evaluationReferenceInputs)

• A typed ReferenceInput model for a single reference entry (context, expected_response, assertions, expected_trajectory), with camelCase aliases matching the wire format and an expected_response_text convenience property.
• A reference_inputs: List[ReferenceInput] field on EvaluatorInput, populated by the decorator from event["evaluationReferenceInputs"].
• ReferenceInput is exported from both custom_code_based_evaluators and the top-level evaluation package.

Evaluator identity (evaluatorId / evaluatorName)

• Optional evaluator_id / evaluator_name fields on EvaluatorInput, mapped by the decorator. Lets an evaluator log or branch on which evaluator config invoked it.

Usage

from bedrock_agentcore.evaluation import (
    custom_code_based_evaluator, EvaluatorInput, EvaluatorOutput,
)

@custom_code_based_evaluator()
def exact_match(inp: EvaluatorInput, context) -> EvaluatorOutput:
    expected = next(
        (r.expected_response_text for r in inp.reference_inputs if r.expected_response_text),
        None,
    )
    if expected is None:
        return EvaluatorOutput(label="Skip", explanation="No ground truth provided")
    ...

Notes

• Backward compatible — all new fields default to empty/None; existing evaluators are unaffected.
• expected_response is modeled as an object ({"text": "..."}), not a bare string, matching the contract. The expected_response_text property encodes this so callers don't misread it.
• ReferenceInput uses extra="allow", so unknown/future keys in a reference entry are preserved (reachable via ref.model_extra) instead of dropped.
• Naming: the package already exports ReferenceInputs (plural, client-side ground-truth input). This adds ReferenceInput (singular, the parsed evaluator-side entry) — opposite ends of the same contract.

Testing

uv run pytest tests/bedrock_agentcore/evaluation/custom_code_based_evaluators/ — 31 passed. New tests cover decorator pass-through (reference inputs + evaluator id/name), default-empty/None, alias coercion, extra-key preservation, and the expected_response_text property. ruff check and ruff format --check clean.

[github-actions]

✅ No Breaking Changes Detected

No public API breaking changes found in this PR.

[agentcore-devx-automation]

Claude Security Review: the review did not analyze this PR (model took 0 turns). See the run for details; a later push or re-run is needed.

[agentcore-devx-automation]

Claude Security Review: the review did not analyze this PR (model took 0 turns). See the run for details; a later push or re-run is needed.

[agentcore-devx-automation]

Claude Security Review: the review did not analyze this PR (model took 0 turns). See the run for details; a later push or re-run is needed.