fix kl mismatch by resetting prefix cache

[hallerite]

After bumping verifiers in #1572, the kl mismatch increased, because Verifiers stopped returning prompt logprobs in PrimeIntellect-ai/verifiers#666 and vLLM defaults to prefix caching in that case, which needs to be reset after weight updates for correctness.

---

Note

Ensures KV cache invalidation when model weights/adapters change to prevent stale-prefix usage.

• Resets prefix cache after update_weights and reload_weights RPCs in server.py
• Adds POST /load_lora_adapter server endpoint wrapping vLLM’s loader; on success resets prefix cache and returns {status: ok}; propagates error responses
• Client now calls /load_lora_adapter (was /v1/load_lora_adapter); updated docstrings/comments
• Unit tests updated to reflect new endpoint and retry behavior

^{Written by Cursor Bugbot for commit 17539f9. This will update automatically on new commits. Configure here.}

[Jackmin801]

Good catch! Can we modify the /update_weights and /reload_weights to do the reset instead of having a new path?

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[Jackmin801]

can we use the basemodel dataclass definition so the swagger docs are correct? its useful for debug sometimes

[Jackmin801]

nice! lgtm