[Doc] Add CLAUDE.md / AGENTS.md with contribution rules for AI agents

[vmoens]

Summary

• Adds CLAUDE.md documenting TorchRL's house rules for LLM-driven contributions, sitting on top of CONTRIBUTING.md.
• Adds AGENTS.md as a symlink to CLAUDE.md so Codex and Claude see the same rules and they cannot drift.

The rules cover:

• Imports: no local imports except for optional deps (gated by _has_<name> = importlib.util.find_spec(...)) or genuine circular imports (try TYPE_CHECKING first); no wildcard imports; from __future__ import annotations at the top of new files.
• Cross-version compatibility: use torchrl.implement_for (from pyvers) instead of hand-rolled version branching.
• Logging / timing: use torchrl_logger instead of print; use torchrl.timeit instead of ad-hoc time.time().
• TensorDict-first APIs and the _AcceptedKeys pattern for new objectives.
• torch.compile / cudagraphs friendliness (encouraged, not mandatory): prefer torch.where over Python control flow, avoid data-dependent shapes / .item() on hot paths.
• Tests: every new public class needs tests; extend existing test files rather than creating new ones; new algorithms must run in the sota-implementations CI.
• Docs: every new public class is referenced in docs/source/reference/*.rst; Sphinx-style docstrings with runnable >>> examples; arXiv refs for paper-based features; no emojis.
• Tutorials for headline features, structured Sphinx-first with "What you will learn" / "Conclusion" / "Further reading" sections, using # prose comments rather than print(...).
• Benchmarks for performance-relevant changes.
• SOTA: new algos ship a Hydra script under sota-implementations/<algo>/ plus a sota-check/ entry.
• Type hints on new public signatures (documentary, but must be accurate).
• Deprecation policy: two-release schedule (deprecate in 0.X, default change in 0.X+1, removal in 0.X+2), using DeprecationWarning / FutureWarning with the target version named explicitly in the message.
• PR labels: canonical [Tag] set documented ([Algorithm] [BE] [BugFix] [CI] [Deprecation] [Doc] [Feature] [Minor] [Performance] [Quality] [Refactor] [Test] [Versioning]).

No squash-commit requirement.

Test plan

• Render CLAUDE.md on GitHub and confirm the formatting reads well.
• Confirm AGENTS.md resolves to CLAUDE.md (GitHub renders symlinks as a link to the target).
• Sanity-check that the rules match current repo conventions (imports, logger, timeit, _AcceptedKeys, implement_for, label set in recent git log).

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/rl/3686

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• Rolling out OSDC (ARC) runners on pull & trunk workflows in PyTorch main

❌ 2 Awaiting Approval, 2 New Failures, 13 Pending

As of commit 5c6aa62 with merge base e479313 ():

AWAITING APPROVAL - The following workflows need approval before CI can run:

• Libs Tests on Linux (gh)
• Tutorials Tests on Linux (gh)

NEW FAILURES - The following jobs have failed:

• Push Binary Nightly / build-wheel-unix (macos, macos-latest, 3.10, cp310-cp310, cpu, cpu) (gh)
  Process completed with exit code 1.
• Push Binary Nightly / test-wheel-unix (macos, macos-latest, 3.10, cp310-cp310, cpu, cpu) (gh)
  Unable to download artifact(s): Artifact not found for name: torchrl-macos-3.10_cpu.whl

This comment was automatically generated by Dr. CI and updates every 15 minutes.