perf(managers): continue reducing update_state cost (#1316)

[TATP-233]

Summary

This integration PR completes roadmap #1316 through three bounded child PRs:

• perf(motion): fuse command metrics in numba #1320 / perf: 降低 command_manager.compute 的 MotionCommand 开销 #1317: replace MotionCommand metric/update NumPy expressions with one prewarmed parallel Numba implementation
• perf(motion): fuse relative transforms in numba #1321 / perf: 降低 command_manager.post_compute 的 relative transform 开销 #1318: fuse relative/body-frame transforms into shared Numba kernels and remove the replaced motion NumPy modules/helpers
• perf(observations): reduce batch pipeline copies (#1319) #1322 / perf: 降低 observation_manager.compute 的批量 pipeline 开销 #1319: reduce observation full-batch copies, finite scans, noise temporaries, and sanitize copies without adding an unsuitable second dispatch path

The two arithmetic replacements are Numba-only in production; NumPy remains only in independent test oracles. The observation stage contains RNG, ownership copies, finite scanning, and concatenation rather than a standalone arithmetic loop, so it was optimized directly instead of maintaining a nominal Numba path. Motrix, public contracts, runner/lifecycle, and config ownership are unchanged.

Closes #1316
Milestone: M2 - Manager update_state throughput

Branch contract

• Declared base: dev/issue-1304-motion-numba-body-state@9cca010d2c97910a6704a3da80b0d7f347103591
• Integration head: dev/issue-1316-update-state-numba@d09d05aa39123ec1001210500e61b6539cb22caf
• Child heads were created from the latest integration branch and merged in order.
• This PR does not target, checkout, push, merge, or rewrite main or dev/issue-1042-manager-based-api.

Direct cumulative A/B

Fresh interleaved comparison of exact roadmap base 9cca010d against the final tree: g1_motion_tracking, 8192 envs, warmup 10, measure 100, fixed action/env seeds, three repetitions on the same host. Phase values are the median of three per-run medians; instrumented manager values are the median of three per-run means.

Backend	Metric	Roadmap base (ms)	Final (ms)	Delta
MuJoCo	command_manager.compute	7.364839	2.113478	-71.30%
MuJoCo	command_manager.post_compute	4.332488	0.710599	-83.60%
MuJoCo	observation_manager.compute	4.431236	4.199805	-5.22%
MuJoCo	update_state	19.417052	10.322279	-46.84%
MuJoCo	step_core	52.925145	52.876745	-0.09%
MuJoCo	env_step_total	84.086522	76.055615	-9.55%
MJWarp	command_manager.compute	6.747341	1.753317	-74.01%
MJWarp	command_manager.post_compute	4.262291	0.586890	-86.23%
MJWarp	observation_manager.compute	3.811295	3.584991	-5.94%
MJWarp	update_state	17.762506	8.971272	-49.49%
MJWarp	step_core	14.396119	14.419870	+0.16%
MJWarp	env_step_total	49.853370	40.943064	-17.87%

The isolated #1319 interleaved A/B, using its immediate base 1d18060e, measured observation_manager.compute at -7.24% MuJoCo and -7.93% MJWarp. The direct roadmap-wide observation deltas above are smaller because all three stages are initialized and sampled together; both measurements agree on direction without claiming sub-noise end-to-end effects for #1319 alone.

MJWarp PCIe accounting

Metric	Roadmap base (ms)	Final (ms)	Delta
backend_physics_ms	13.604566	13.669142	+0.47%
backend_host_cache_refresh_ms	0.658934	0.659782	+0.13%
backend_control_upload_ms	0.103609	0.097939	-5.47%

The approximately 0.66 ms GPU-to-host copy and synchronization happen inside backend.step(). They are therefore already included in step_core_ms and env_step_total_ms, while remaining outside update_state; the table reports them separately and does not double-count them.

Remaining bottlenecks

At the final head, the largest measured update_state partition is still observation_manager.compute (MuJoCo/MJWarp 4.20/3.58 ms), followed by complete reward+termination term calls (2.94/2.62 ms) and command_manager.compute (2.11/1.75 ms). command_manager.post_compute is now 0.71/0.59 ms. For the full environment step, backend physics remains dominant; on MJWarp, the stable 0.66 ms D2H barrier is part of the real collector cost.

Hardware: AMD Ryzen 9 9950X3D2 (16 cores / 32 threads), NVIDIA GeForce RTX 4090 48 GiB, Linux 7.0.0-30-generic, NumPy 2.4.4, Numba 0.67.0, 32 Numba threads.

Validation

• every child final head passed local make test-all before its PR was created
• exact integration head d09d05aa passed make test-all: 2309 passed, 28 skipped, 281 deselected, 1 xfailed; mypy, pyright, Ruff, and benchmark import smoke passed
• parity covers full batch, partial-reset rows, untouched rows, dtype/shape, quaternion semantics, output ownership, fixed-seed noise, delay/history, and NaN policies
• working tree is clean

Per #1313 governance this is a non-main integration PR into the declared non-protected base. Remote CI is intentionally not scheduled or awaited; the complete gate is the recorded local make test-all.