harbor_env: forward purpose, provider and eval_sampling from HarborSessionFactory

[jayzuccarelli]

HarborSessionFactory (where #1276 points opencode_env users) can't pass purpose, provider or eval_sampling, although HarborEnv.run_rollout and the server's run_rollout tool accept all three. So a TRL loop-owning run can't ask for purpose="train", which the provider qualification guide recommends.

That matters most for training: a vLLM started without --return-tokens-as-token-ids still runs every rollout, as eval, and fetch_proxy_trace() hands the trainer an empty list per rollout with only a warning. With purpose="train" the server refuses the rollout before creating a sandbox.

The factory now takes the three kwargs and passes them through. It validates them at construction, like sampling, so a bad value fails once rather than on every rollout. Defaults are unchanged.

Tests: new cases in tests/envs/test_harbor_session_factory.py fail on main (TypeError on the new kwargs) and pass with this change; the harbor and capture suites pass (606 passed, 3 skipped). ruff and usort clean on the touched files.

Prepared with AI assistance (Claude Code); I reviewed the change and ran the tests.

---

Note

Low Risk
Additive API with constructor validation and passthrough to existing rollout parameters; no changes to auth or core training logic beyond enabling explicit train/eval intent.

Overview
HarborSessionFactory and HarborSession now accept provider, purpose, and eval_sampling and pass them through to HarborEnv.run_rollout, matching what the harbor server already supports. Defaults stay openai, auto, and no eval sampling.

Factory construction validates these up front (like sampling): purpose must be auto, eval, or train; eval cannot combine with training sampling; eval_sampling requires purpose="eval" and is checked via evaluation_sampling.

This lets TRL loop-owning runs request purpose="train" so rollouts fail immediately when the engine cannot return token ids, instead of running as eval and yielding empty traces.

Tests cover forwarding purpose, provider, and eval_sampling, plus parametrized rejection of invalid combinations at factory init.

^{Reviewed by Cursor Bugbot for commit f567900. Bugbot is set up for automated code reviews on this repo. Configure here.}