fix: enum values in input schema

[cotovanu-cristian]

Summary

This PR fixes prompt interpolation for enum-backed input fields so the LLM receives the actual enum value instead of the enum member name. In the failing case, risk_tolerance=low was being rendered as DynamicEnum.KEY_0, which
changed the prompt semantics and led to invalid model behavior.

Changes

• Update prompt argument serialization to unwrap enum-like values before converting them to strings
• Preserve existing JSON serialization behavior for lists, dicts, and booleans
• Add a unit test covering string-enum interpolation

Why

Agent prompts must reflect the validated input payload exactly. Passing enum member names instead of their values breaks downstream reasoning and can cause the model to branch incorrectly.