[aknight-ld]

Python and JS SDKs had a subtle disagreement in how outputs were formatted to send to judges. Making the decision to spec these outputs as json objects, and adjusting the code to match more closely.

---

Note

Overview
Aligns how generation output is shown to judges when it is not a plain string: response_to_evaluate and the output segment of message_history now go through tool_result_string, so dict/structured results become JSON text (same rule as tool results) instead of Python’s default string form.

Offline evaluation wiring is updated so judge handlers receive the formatted response_to_evaluate variable rather than the raw row output, keeping the judge user_input argument consistent with template variables.

Tests cover end-to-end eval runs and build_message_history (including non-ASCII and empty dict output).

^{Reviewed by Cursor Bugbot for commit f36a14b. Bugbot is set up for automated code reviews on this repo. Configure here.}