Fix Pydantic JSON types in strict evidence serialization

[ereinach]

Rationale

Typed Pydantic inputs containing JSON-specific values such as HttpUrl fail before RLM execution with Strict evidence serialization failed for run.started. This violates the established plain-data normalization and strict-evidence contract: supported Pydantic inputs must become JSON-serializable evidence.

Reproduction: emit a run.started evidence event whose inputs contain a Pydantic model with an HttpUrl. to_plain_data() used model_dump(mode="python"), which preserved the HttpUrl; the subsequent json.dumps() raised TypeError.

The root cause is that Pydantic v2 normalization selected Python mode even though its result is used for JSON transport and evidence.

Summary

• Serialize Pydantic v2 models with model_dump(mode="json").
• Add a behavior-level regression covering an HttpUrl nested in run inputs.
• Preserve the existing fallback for model-dump-compatible objects that do not accept a mode.

Test Plan

• uv run pytest tests/test_small_kernel.py::test_strict_evidence_serializes_pydantic_json_types -q
• uv run pytest tests/test_small_kernel.py -m "not sbx" -q (101 passed, 4 optional SBX tests deselected)
• uv run ruff check src/predict_rlm/serialization.py tests/test_small_kernel.py