[luyuzhe111]

Summary

sagemaker_rft_handler currently reports dictionary results without a reward as ready without submitting a reward. Invalid reward values can also trigger ready before reward submission fails.

Validate rewards before reporting success. Accept finite integers and floats, or non-empty lists of those values; reject missing values, None, booleans, non-numeric values, and non-finite floats. Invalid rewards use the existing report_error path to report failed with reward 0. The handler's return value, explicit error/skipped handling, public signatures, and feedback exception policy remain unchanged.

This tightens automatic reporting: a decorated handler that manually reports feedback but returns a dictionary without a valid reward will now also attempt to report failed.

Validation

• RFT unit tests: 36 parameterized cases passed across synchronous and asynchronous handlers. Tests run the real decorator and feedback client, mocking only HTTP submission to verify the complete feedback request sequence and unchanged return values.
• Black and flake8 passed for both changed Python files.

[luyuzhe111]

cc @992X @SifeiLi for review. issue identified by @qingliz in awslabs/agentcore-rl-toolkit#209