Fix cache salt forwarding in train client

[eligotts]

Summary

Forward cache_salt through Renderers’ dedicated generate argument instead of leaving it inside sampling_params, where it does not reach vLLM.

This preserves policy-version prefix-cache isolation for pipeline RL: existing rollouts retain their starting cache namespace, while fresh rollouts miss KV cached under older weights.

Validation

• uv run ruff check verifiers/v1/clients/train.py
• Commit hooks: Ruff check, Ruff format, and ty

Note

Pass cache_salt as separate argument in TrainClient.get_response

Extracts cache_salt from sampling_params via pop and forwards it as a distinct argument to the generate call in train.py. sampling_params no longer carries cache_salt when forwarded to generate.

^{Macroscope summarized 7367a4a.}

---

Note

Low Risk
Single-path wiring change in the train client; no auth or data-model changes, aligned with the legacy renderer client pattern.

Overview
Fixes cache_salt not reaching vLLM when the train client calls renderers.client.generate.

TrainClient.get_response now pops cache_salt from sampling_params (same pattern as chat_template_kwargs) and passes it as a dedicated cache_salt argument to generate, instead of leaving it inside sampling_params where it was ignored.

This restores prefix-cache namespace isolation for pipeline RL rollouts that key KV cache on policy version.

^{Reviewed by Cursor Bugbot for commit 7367a4a. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Approved at 7367a4a

Macroscope's review found this PR approvable — This is a focused, single-file bug fix that forwards an existing cache-salt value through the renderer's intended dedicated argument. Requests without a cache salt retain their prior behavior, while configured train rollouts regain the intended prefix-cache isolation.

^{You can add or adjust custom eligibility rules. Learn more.}