refactor: make openadapt-ml a leaf; break ml<->evals import cycle

[abrichr]

Summary

Phase 1 of the openadapt evals→ml refactor. Makes openadapt-ml a clean model-library leaf by removing every module-level import of openadapt-evals.

Depends on openadapt-types PR #5 (adds openadapt_types.benchmark).

Fork A — Benchmark* vocabulary from openadapt-types

The one hard, unconditional module-level ml → evals import was openadapt_ml/benchmarks/agent.py importing BenchmarkAction/BenchmarkAgent/BenchmarkObservation/BenchmarkTask from openadapt_evals. These types now live in openadapt-types; agent.py (and the lazy BenchmarkAction imports in trainer.py / runner.py) import them from there.

Fork B — thin RolloutEnv Protocol in openadapt-ml

New openadapt_ml/training/grpo/rollout_env.py defines a minimal RolloutEnv Protocol (reset/step/observe/collect_rollout/screen_size) — the surface the GRPO trainer/collector actually drive. The concrete WAA envs/adapters (RLEnvironment, WAALiveAdapter, WAADesktopEnv) stay in openadapt-evals and implement it structurally. ml depends on the interface; evals provides the implementation.

Cleanup — remaining module-level evals imports made lazy

Converted the guarded module-level openadapt-evals imports to lazy in-function imports:

• rollout_collector.py — env/adapter imported inside __init__ (TYPE_CHECKING for annotations)
• verl_backend.py — WAADesktopEnv behind _load_waa_desktop_env()
• trainer.py — TaskConfig inside _load_task_configs

Verification

$ grep -rn "import.*openadapt_evals\|from openadapt_evals" openadapt_ml/ --include=*.py

All remaining hits are lazy (in-function), TYPE_CHECKING-only, or docstrings — no module-level ml → evals import remains. import openadapt_ml.benchmarks.agent now succeeds with openadapt-evals not installed.

Tests: pytest tests/ --ignore=tests/integration → 451 passed, 5 skipped. (tests/integration requires a live openadapt-evals install and is excluded in CI, as before.)

⚠️ Release sequencing (IMPORTANT — do not merge before types)

1. openadapt-types PR fix: lower Python requirement from 3.12 to 3.10 #5 merges → cut an openadapt-types release (adds openadapt_types.benchmark).
2. Update the pin in pyproject.toml here: currently openadapt-types>=0.2.0 with a local editable [tool.uv.sources] entry (../openadapt-types) so imports resolve pre-release. Replace the editable source with the published version once released.
3. Then merge this PR.

Pre-existing junk (left untouched)

This working tree has pre-existing untracked junk NOT part of this change and deliberately not committed: .git-backup-*.bundle, test_artifacts/, demo .txt files under experiments/waa_demo/annotated_demos/, and a self-referential openadapt-ml symlink at repo root (the symlink causes recursive test collection if you run bare pytest from the root — run pytest tests/ as CI does).

🤖 Generated with Claude Code

https://claude.ai/code/session_01CKrVJJy5jWVCkXAqgUqtqZ