test: add renderer client tests for chat_template_kwargs materialization

[mvanhorn]

Summary

Adds renderer-client test coverage proving that chat_template_kwargs are materialized into the rendered prompt and that omitting them falls back to each model's chat-template default rather than a hardcoded value.

Why this matters

Issue #2572 asks for tests confirming that the renderer path in src/prime_rl/utils/client.py respects model-specific defaults and that enabling renderer-based tokenization did not silently break other model families. The cases called out are reasoning_effort for gpt-oss and enable_thinking for qwen, nemotron, and laguna; there was no prior coverage of this behavior.

Testing

tests/unit/utils/test_client.py gains a parametrized class covering: reasoning_effort passed for gpt-oss appearing verbatim in the rendered prompt, enable_thinking=false suppressing reasoning markup for qwen/nemotron/laguna, omitting enable_thinking resolving to the tokenizer chat-template default, an unsupported kwarg surfacing a clear error, and a non-renderer path left unaffected. The change is test-only; src/prime_rl/utils/client.py is untouched. Covered by the new tests in this PR; full suite runs in CI.

Fixes #2572

---

Note

Low Risk
Test-only changes with no runtime or security impact.

Overview
Adds unit tests in tests/unit/utils/test_client.py that exercise the renderer client path (via a stub RendererClient and prompt-capturing fake renderer) to lock in how sampling_args.extra_body.chat_template_kwargs affect rendering.

reasoning_effort for gpt-oss is expected to show up in the rendered prompt and not be forwarded on the native generate sampling_params. enable_thinking is covered for Qwen 3.5, Nemotron-3, and Laguna: false suppresses reasoning markup; when omitted, behavior follows per-model defaults (parametrized). Passing an unsupported kwarg (e.g. enable_thinking on gpt-oss) must raise a clear ValidationError.

A separate test asserts setup_clients does not attach renderer_config / renderer_model_name to the default OpenAI chat completions client when only renderer-related args are passed—guarding the non-renderer path.

No production code changes; this is regression coverage for issue #2572.

^{Reviewed by Cursor Bugbot for commit d1c2edf. Bugbot is set up for automated code reviews on this repo. Configure here.}