[jhnwu3]

…nner

• Pluggable LLM backends (backends.py): Gemini and OpenAI-compatible servers (vLLM), selectable per student; google-genai is now an optional extra. Non-retryable API errors (e.g. context overflow) end the student's attempt instead of crashing the run; Qwen-style reasoning blocks are stripped.
• TaskEnvironment seam (environments.py) threaded through the learning cycle and final evaluation; the upstream DataFrame path is unchanged.
• PyHealth environments (pyhealth_env/): students write solution(train_dataset, val_dataset, device) -> model; each attempt trains in a fresh subprocess on a GPU slot (GpuPool) and is scored on a held-out patient-level split. Built-in MIMIC-III/IV and eICU specs, reuse of cached SampleDatasets, prompts with a real-batch schema table plus PyHealth Trainer and custom BaseModel templates.
• Headless runner (scripts/run_classroom.py, run.py, configs/): YAML configs with dotted overrides; notebook/Colab/Plotly code removed.
• pyproject: repro extra pinning the classroom env used for reported results, separate vllm-server extra, explicit package discovery.
• README: two-env setup, install pitfalls, vLLM recipes for RTX 2080 Ti and A6000, and reference MIMIC-IV mortality results (Qwen3.5-9B-AWQ students: test AUROC 0.785-0.792 vs 0.781 for the PyHealth RNN baseline).

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.