test: add LiteLLM schema selection runtime coverage

[rlippmann]

Summary

Adds Tier 2 runtime-boundary coverage for the existing Python LiteLLM response_format schema-selection example.

The new tests exercise the real litellm.completion(...) call path against a local OpenAI-compatible stub server and assert on the actual HTTP payload received by the stub.

Why

The existing schema-selection coverage verified host-side selection behavior, but did not prove that the selected response_format crossed a real LiteLLM runtime boundary.

This change validates that schema selection affects the actual LiteLLM request payload without requiring an OpenAI, Anthropic, or other frontier-model key.

Changes

• Added test_litellm_runtime_sends_selected_response_format
• Added test_litellm_runtime_omits_response_format_when_unselected
• Kept example behavior unchanged

Scope

This PR does not:

• create a new example
• change schema-selection behavior
• call a live model provider
• introduce model-derived state mutation

Notes

The tests use an ephemeral localhost OpenAI-compatible stub server. They verify:

• selected schemas are serialized into the LiteLLM outbound request as response_format
• unselected schemas omit response_format
• the example still parses LiteLLM responses successfully

Tier 2 Coverage Impact

This closes the schema-selection runtime-boundary gap identified during the Tier 2 audit.