[wlgys8]

Summary

• Move CUDA/ROCm extras, version pins and extra-gated PyTorch index routing into the installable motrix-rl package.
• Forward workspace GPU extras to motrix-rl while preserving install.sh entry points.
• Per reviewer request, do not retain heavyweight resolver integration tests; source consumption, install plans and wheel metadata were checked as one-time acceptance.
• Document external consumer conflicts and the distinction between source-based uv installation and wheel/PyPI installation.

Fixes #29

Validation

• One-time GPU resolver/wheel acceptance — 11 checks passed before removing the temporary integration tests. Retained version tests: python3 -m pytest tests/test_workspace_versions.py -q
• Verified resolver tests with uv 0.9.8 and CI version 0.9.14 before rebase
• uv lock --check
• prek run --all-files
• git diff --check
• Full training suite / actual GPU runtime testing (not run; dependency-only change)

Compatibility and limitations

• Existing workspace install.sh --gpu cuda/rocm entry points remain unchanged.
• Default installs continue resolving Torch from PyPI; GPU index routing is activated only by the corresponding extra.
• External Git consumers exposing both optional profiles must declare root forwarding and package-qualified conflict pairs; selecting only one profile needs no extra conflict configuration.
• Wheel metadata does not preserve uv index routing; pip and wheel consumers must configure the PyTorch index explicitly.
• uv 0.9.8/0.9.14 export loses GPU selection through forwarding extras. Documented limitation: use uv sync; one-time acceptance inspected isolated sync --dry-run plans rather than exported requirements.
• No GPU wheels installed and no existing runtime environment modified by validation.

Target branch

Targets upstream main; rebased onto the merged Python 3.11 support PR so this PR includes only issue #29 changes.

[Copilot]

🔵 Needs a closer look

Correctness hinges on uv's workspace resolution semantics and a large auto-generated lockfile that cannot be fully validated in this environment (no uv/network), affecting installation for all GPU and external source consumers.

No open issues 🎉

What changed in this PR

This PR implements issue #29 by relocating the GPU wheel dependency profiles (CUDA/ROCm torch pins and PyTorch index routing) from the non-installable workspace root into the installable motrix_rl package. The goal is to let external projects that consume motrix_rl via a Git/source dependency inherit the pinned torch versions and extra-gated index routing automatically, instead of re-copying ~40 lines of configuration. The workspace root now forwards its cuda/rocm extras to motrix-rl[cuda]/motrix-rl[rocm], so repository-internal install.sh entry points are preserved.

Changes:

• Move cuda/rocm extras, the pytorch-cu128/pytorch-rocm72 explicit indexes, extra-gated tool.uv.sources torch routing, and the cuda/rocm conflict into motrix_rl/pyproject.toml; root forwards via cuda = ["motrix-rl[cuda]==0.4.0b1"] / rocm = [...].
• Add extra-gating to the torchvision/torchaudio/triton-rocm sources (previously ungated), tightening default-index resolution.
• Document GPU install profiles and external-consumer conflict/source requirements in motrix_rl/README.md, and expand testing-layering guidance in AGENTS.md.
• Regenerate uv.lock to reflect the new extra-9-motrix-rl-cuda/-rocm extras.

File	Description
motrix_rl/​pyproject.toml	Adds cuda/rocm optional-dependencies, pytorch indexes, extra-gated sources, and member-local conflict.
pyproject.toml	Replaces root GPU extras/indexes/sources with thin forwarding extras to motrix-rl; keeps the mutually-exclusive conflict.
motrix_rl/​README.md	Documents GPU profiles, source-vs-wheel install behavior, and external Git-consumer conflict declaration.
AGENTS.md	Expands testing guidelines on long-term tests vs one-time acceptance and test layering.
uv.lock	Auto-regenerated lock reflecting the new per-package extras and markers.

I verified: version pins are consistent at 0.4.0b1; tests/test_workspace_versions.py self-dependency/pin logic still passes with the forwarding extras; no stale pytorch-cu128/pytorch-rocm72 references remain elsewhere; install.sh --gpu cuda/rocm and CI uv sync --frozen --extra cuda paths continue to route through the root forwarding extras. I found no objective line-level defects in the hand-edited files.

_{Review effort: Balanced}