perf(mjwarp): reduce g1 motion tracking reset latency

[TATP-233]

Summary

• Scale the MJWarp reset scratch allocation on the cold path (128/256/512 worlds by batch size, capped at 512) so the 8192-env motion-tracking workload stays on the bounded sparse-forward route for its usual ~250 reset rows.
• Add MJWarp selected-row world body pose/velocity getters, matching the existing SimBackend row contract and avoiding full-batch cache materialization during MotionCommand partial reset.
• Preserve the existing full-forward path for oversized resets and all eager/graph-disabled fallback behavior; MuJoCo, task config, runner, learner, and lifecycle code are unchanged.

Linked Work

• Issue: Closes perf(reset): 定位 MJWarp 在 SAC g1_motion_tracking/8192 的 collector reset 瓶颈 #1325
• Declared integration base: dev/issue-1304-motion-numba-body-state
• Implementation branch: perf/issue-1325-mjwarp-reset
• Base commit: 4fa6631a3c9c465db86cd671275319168ce36fe0

Scope and governance

• Owner layer: src/unilab/base/backend/mjwarp/ and focused backend tests only.
• No new public contract, task/config change, runner/lifecycle path, or other backend change.
• The PR is intentionally not configured for auto-merge; maintainer review and manual merge are required.
• This PR targets the declared dev integration branch, not main; per the governance update in PR docs: streamline roadmap and PR governance #1313, the local make test-all result is the child-PR gate.

Validation

• Focused host-cache/CUDA-graph/smoke tests
• Real CUDA MJWarp slow tests and short differential tests
• make test-all
• git diff --check

Commands actually run:

uv run pytest -q tests/base/test_mjwarp_host_cache.py tests/base/test_mjwarp_cuda_graph.py tests/base/test_sim_backend_smoke.py
uv run --extra mjwarp pytest -q -m slow tests/base/test_mjwarp_backend.py tests/base/test_mjwarp_cuda_graph.py tests/base/test_mjwarp_host_cache.py tests/base/test_mjwarp_differential.py tests/base/test_mjwarp_capabilities.py tests/base/test_backend_conformance.py
make test-all
git diff --check

Results:

• Focused tests: 40 passed.
• MJWarp CUDA/differential set: 19 passed, 45 deselected.
• make test-all: 2316 passed, 28 skipped, 281 deselected, 1 xfailed; benchmark smoke 33/34 module-mode and 34/35 script-mode (the one skip in each is platform-optional MLX).
• Static checks inside make test-all: ruff, mypy (247 source files), and pyright all passed.

Performance evidence

Same-host SAC g1_motion_tracking, 8192 envs, 10 warmup + 100 measured vector steps, using benchmark_offpolicy_collector_active.py:

Metric	Baseline (issue investigation)	This branch
collector active throughput	184,852 env/s	229,587 env/s
env step	42.553 ms	33.995 ms
reset_done	14.501 ms	5.859 ms
reset_done_reset_call	14.348 ms	5.708 ms
reset rows / vector step	251.59	251.4
MJWarp set_state	5.995 ms	1.572 ms
MJWarp reset forward sub-item	5.206 ms	1.276 ms

The final run selected a 512-world scratch data for the 8192 batch; the measured reset-row population remained below that bound. MuJoCo source behavior is unchanged; in the paired final run its reset_done was 6.643 ms and throughput 85,497 env/s (within run-to-run variation of the investigation baseline).

Impact

• Backend impact: MJWarp only; MuJoCo/Motrix code paths are unchanged.
• Platform impact: Linux/CUDA MJWarp path; CPU-only and graph-disabled paths retain their existing fallback.
• Training effect expected: no change to observations, reset semantics, action/reward dimensions, RNG, or sim2sim configuration; only reset latency should improve for bounded sparse resets.

Checklist

• Added/updated focused tests for capacity routing, selected-row parity, and no full-batch getter fallback.
• No documentation or configuration change required.
• Linked the driving issue.
• Noted oversized/graph-disabled fallback and manual-merge follow-up.