test: align contract coverage and isolate integration suites

[wlgys8]

Summary

• Apply the AGENTS.md testing policy across environment, deployment and RL suites: remove adjustable task-default snapshots, obsolete migration absence assertions, fake-only and duplicate checks.
• Replace implementation-shape assertions with observable behavior: configured randomization bounds, repeated simulator opens, physics forwarding and ONNX clipping parity.
• Reduce heavyweight fixtures: test noise distribution through the production noise kernel; use synthetic motion-library data; test invalid duration at the CLI boundary without repeated full subprocess launches.
• Make metrics backoff regression deterministic with controlled time and synchronization; guarantee sender cleanup. Isolate DDP rendezvous and clean up both workers on failure; remove unbounded drain loops.
• Classify real network, subprocess/distributed, simulator, ONNX and GPU tests as integration while retaining lightweight unit cases. Document test selections in AGENTS.md.
• Fix the stale motrix-deploy runtime version by reading distribution metadata and testing the consistency relationship instead of a hardcoded version.

Validation

• Repository fast suite: 1676 passed, 5 skipped (not slow and not integration and not numba).
• Dedicated Numba/slow CPU suite: 35 passed.
• Integration suite excluding test/test_all_envs.py: 194 passed.
• Additional targeted environment, deployment, provider, CLI and registry suites passed during implementation.
• Metrics backoff regression passed 20 consecutive runs.
• prek run --all-files
• git diff --check
• uv lock --check

Compatibility and limitations

• Training/runtime behavior unchanged except resolving the stale deployment package version from installed distribution metadata.
• No new project dependencies or lockfile changes.
• Validation used Python 3.10 with bounded CPU/Numba threads. Torch tests temporarily used existing system NVIDIA library directories through LD_LIBRARY_PATH because the venv lacks some shared libraries; Torch was not reinstalled.
• Full environment-catalog smoke, actual GPU execution, real DDS hardware and interactive viewers were not validated. Hardware/viewer tests use injected implementations; simulator and ONNX CPU integration tests were executed.
• Based on latest main after PR feat(rl): expose GPU profiles to source consumers #129 merged; this PR contains only the testing cleanup.

[Copilot]

🔵 Needs a closer look

It spans ~40 files and reclassifies many tests out of the default CI suite, a coverage-impacting decision that warrants human review despite the changes being almost entirely test-only.

1 open issue

• Tautological version assertion compares the same metadata value · New

What changed in this PR

This PR applies the AGENTS.md testing policy across the whole workspace. It removes brittle snapshot assertions (adjustable task defaults, migration-absence checks, duplicate/fake-only tests), replaces implementation-shape assertions with observable-behavior checks, isolates heavyweight tests behind @pytest.mark.integration/@pytest.mark.slow, makes the metrics-backoff and DDP tests deterministic, and fixes the stale motrix-deploy version by reading distribution metadata. It is almost entirely test-only; the single production change replaces a hardcoded __version__ with importlib.metadata.version("motrix-deploy").

Changes:

• Reclassify real-network/subprocess/simulator/ONNX/GPU tests as integration, keeping lightweight unit cases in the fast suite, and document the selection split in AGENTS.md.
• Replace snapshot/implementation-shape assertions with behavior- or relationship-based checks (randomization bounds, drain counts, friction forwarding, ONNX clipping parity, noise distribution via the production kernel).
• Harden flaky/heavyweight tests: deterministic metrics backoff via a controlled clock/condition, FileStore-based DDP rendezvous with guaranteed worker cleanup, CUDA-index test without a real GPU, and resolve the deploy version from metadata.

File	Description
AGENTS.md	Documents the fast/numba/integration pytest selections and integration scope.
motrix_deploy/​src/​motrix_deploy/​__init__.py	Reads __version__ from distribution metadata instead of a hardcoded string.
motrix_deploy/​tests/​test_package.py	Updates version test to compare against metadata (now tautological); marks CLI/import tests integration.
motrix_deploy/​tests/​test_config.py	Adds CLI _duration_steps boundary test; drops adjustable-default snapshots.
motrix_deploy/​tests/​test_{policy,artifact}.py	Marks ONNX/CLI inspect tests integration.
motrix_rl/​tests/​test_metrics.py	Deterministic backoff via controlled clock/condition; socketless fixtures; integration markers.
motrix_rl/​tests/​test_runner.py	Marks live-metrics test integration.
tests/​test_cli_config.py	Resolves provider config type via frameworks.get_config_type; drops default snapshots.
tests/​test_bench_duration.py	Marks spawn-worker tests integration.
motrix_rl_builtin/​tests/​*	Isolates CUDA/DDP/IPC/spawn tests; strengthens drain-count and trainer-selection assertions.
motrix_rl_skrl/​tests, motrix_rl_rslrl/​tests	CPU-only mock device; ONNX clipping parity via reference; seed set-compare; integration markers.
motrix_envs/​tests/​*	Noise tested through production batch kernel; randomization bounds from config; removes snapshot/shape assertions.
motrix_env_core/​tests, motrix_env_motrixsim/​tests	Backend-inference tests via real base classes; removes duplicate/abstract-surface tests.
motrix_deploy_mujoco/​*, motrix_deploy_motrixsim/​*, motrix_deploy_unitree/​*, motrix_deploy_tasks/​*	Broad integration marking; behavior-based physics/friction/limit checks replacing monkeypatch call-count snapshots.

_{Review effort: Balanced}