Pass renderer chat template kwargs from sampling

[eligotts]

Summary

• Teach RendererClient to consume sampling_args.extra_body.chat_template_kwargs.
• Remove those kwargs from the /generate sampling params and pass them into renderer pool construction instead.
• Include the kwargs in the shared renderer-pool cache key so different template settings do not collide.

We are extracting renderer template controls from chat_template_kwargs in sampling extra_body, which matches the hosted config path. Longer term, this should move toward typed pydantic config models with explicit renderer-specific template args; for now, renderers validate accepted kwargs and alert on unsupported keys.

Companion PRs

• Prime RL: Use sampling chat template kwargs for renderer RL prime-rl#2605
• Renderers: Add renderer chat template kwargs passthrough renderers#60

Validation

• uv run ruff check verifiers/clients/renderer_client.py tests/test_renderer_client.py
• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 uv run pytest tests/test_renderer_client.py -k "renderer_client_honors_configured_renderer_name or renderer_client_uses_renderer_model_name_override or renderer_client_consumes_sampling_chat_template_kwargs"

Note

Pass chat_template_kwargs from sampling args to renderer pool in RendererClient

• RendererClient.get_native_response now extracts chat_template_kwargs from extra_body in sampling args and passes them to _get_renderer_or_pool, rather than forwarding them downstream in sampling_params.
• Distinct chat_template_kwargs values produce separate renderer pool cache entries, using a deterministic frozen representation as the cache key.
• A _pop_chat_template_kwargs helper removes the key from sampling_params and raises ValueError if the value is present but not a mapping.

^{Macroscope summarized 3443fdf.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This change propagates chat_template_kwargs from sampling parameters to renderer pool creation, enabling new configuration options that affect how renderers process requests. Since this introduces a new runtime behavior path for template kwargs (like enable_thinking), human review is recommended.

^{You can customize Macroscope's approvability policy. Learn more.}