Capture RLM agent patches in state

[rasdani]

Summary

• replace the patch-specific Harness.agent_patch_state_key / ComposableEnv git helpers with a generic Harness.state_collectors lifecycle hook
• add GitPatchCollector, which snapshots the post-setup worktree with git tree plumbing through a temporary index and stores the post-rollout diff in state["agent_patch"]
• wire rlm_harness() to use GitPatchCollector() so research-environments2's rlm_swe composition gets patch capture without duplicating SWE taskset or RLM harness internals

Notes

• Collection still runs before scoring, so SWE rubrics can mutate test files afterward without corrupting the captured agent patch.
• The collector does not create commits, stage in the real index, reset the worktree, or depend on SWE-specific taskset code.

Tests

• uv run pytest tests/test_composable_env.py tests/test_rlm_composable_env.py -q
• uv run ruff check verifiers/envs/experimental/composable/composable_env.py verifiers/envs/experimental/composable/harness.py verifiers/envs/experimental/composable/harnesses/rlm.py verifiers/envs/experimental/composable/state_collectors.py tests/test_composable_env.py tests/test_rlm_composable_env.py
• uv run ruff format --check verifiers/envs/experimental/composable/composable_env.py verifiers/envs/experimental/composable/harness.py verifiers/envs/experimental/composable/harnesses/rlm.py verifiers/envs/experimental/composable/state_collectors.py tests/test_composable_env.py tests/test_rlm_composable_env.py
• pre-push hooks: ruff check, ruff format, Sync AGENTS.md from docs, ty (ci parity)

---

Note

Medium Risk
Adds new lifecycle hooks that execute extra sandbox commands during setup/rollout, which can affect timing and state mutation if collectors misbehave (though failures are logged and ignored). Git-based patch collection depends on repo state and tool availability inside the sandbox, so it may introduce intermittent missing/empty artifacts.

Overview
Adds harness-owned state collectors to ComposableEnv, running optional post_sandbox_setup and post_rollout hooks to persist rollout artifacts into state while swallowing collector failures.

Introduces GitPatchCollector, which snapshots a post-setup git tree using a temporary index and then stores a binary/full-index diff of agent edits into state["agent_patch"] after rollout. Enables this collector by default in rlm_harness, exports it from verifiers.envs.experimental.composable, and updates docs/tests to cover collector execution and patch correctness (including untracked/staged changes without modifying HEAD).

^{Reviewed by Cursor Bugbot for commit cb12ae9. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

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

^{Reviewed by Cursor Bugbot for commit 1476d7d. Configure here.}