Refactor experimental RLM checkout caching

[snimu]

Description

• Introduce commit-keyed experimental RLM checkout caching with rlm_ref
• Add shared experimental utils for git checkout caching and upload locks
• Make stale cached checkouts auto-prune when they are no longer in use

Changes:

• add verifiers.envs.experimental.utils.git_checkout_cache for ref resolution, commit-keyed cache materialization, and stale-checkout pruning
• add verifiers.envs.experimental.utils.file_locks for shared/exclusive path locks during archive creation and cleanup
• switch the experimental RLM harness to use rlm_ref (default main) instead of branch-keyed cache resolution
• keep local_checkout as a validation-only bypass for an existing checkout
• memoize the resolved checkout per harness instance so one run stays pinned to one resolved commit
• update ComposableEnv archive creation to hold a shared lock while reading local upload dirs
• update the focused experimental RLM/composable tests to cover the new cache and locking behavior

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test improvement

Testing

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

Note

Medium Risk
Changes how git checkouts are mirrored, materialized, and pruned, and adds filesystem locking around uploads; mistakes could cause flaky caching behavior or unintended deletion of cached worktrees. Scope is limited to experimental harness/utilities but touches subprocess/git and concurrency-sensitive paths.

Overview
Refactors experimental RLM checkout resolution to use a new shared verifiers.envs.experimental.utils package, including generic file_locks and a git_checkout_cache that maintains a repo mirror plus per-commit worktrees and supports branch/tag/full-SHA refs.

Updates the RLM harness API from rlm_branch to rlm_ref, treats local_checkout as a validation-only bypass, memoizes the resolved checkout per harness instance, and prunes stale cached worktrees when they are not locked for upload.

Adds upload-time shared locking in ComposableEnv._build_dir_archive to avoid pruning/races while archiving local directories, and expands test_rlm_composable_env.py to cover ref-to-commit materialization, per-instance pinning vs. new instance updates, lock-protected stale checkout retention, and GH token redaction on clone errors.

^{Reviewed by Cursor Bugbot for commit 6f0d86b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 3 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit b0ec724. Configure here.}