test: stabilize benchmark timing contract

[TATP-233]

Summary

• replace a flaky ordering assertion between two independently sampled wall-clock timings in the G1 motion-tracking Numba benchmark test
• assert the actual record contract: total time is the full update-state measurement, the NumPy/Numba total fields match their update-state fields, and both reward/termination samples remain positive
• refresh and promote Phase 7 evidence because the final gate binds the complete tracked source tree

Part of #705
Addresses #855

Real CI Root Cause

GitHub Actions run 30738994557 passed Ruff, mypy, pyright, benchmark smoke, and shards A/B/C/D/F/H. Shard E failed at:

tests/benchmark/test_g1_motion_tracking_numba_benchmark.py::test_g1_motion_tracking_numba_benchmark_builds_records_and_matches_numpy

The test compared numpy_total_ms >= numpy_reward_termination_ms, but those values are produced by separate time_call invocations. On the shared runner, the one-sample reward/termination timing was 1.647 ms while the independently measured full update-state timing was 1.107 ms. That ordering is not a stable wall-clock contract and does not indicate incorrect benchmark output.

The implementation defines numpy_total_ms = numpy_update_state_mean and numba_total_ms = numba_update_state_mean. The revised test checks those exact definitions and keeps parity, positive timings, record identity, thread, and speedup assertions intact.

Validation

• CI-equivalent shard E: 119 passed, 1 skipped, 5 deselected
• focused regression: 1 passed
• LD_LIBRARY_PATH=/tmp/unilab-nvml-570.133.07/usr/lib/x86_64-linux-gnu make test-all: 2350 passed, 26 skipped, 449 deselected, 1 xfailed; coverage 68%; Ruff, mypy, pyright, and benchmark smoke passed
• benchmark smoke: module mode 36/37 and script mode 37/38; only optional MLX was skipped
• Phase 0-7 standalone gates, baseline, thresholds, DR inventory, claim inventory, backend isolation, support, task rollout, legacy retirement, workflow trigger, and CI shard audits passed
• refreshed Phase 7 source: 51fc5014222306dee30a48ddbe2c19b7a1fafe8f
• refreshed Phase 7 artifact SHA256: 6f3a8576f60efa77a3191d50cd10c78aa4adbfb3d31ebe8cbfbc91e561bcd5de
• Phase 7: 12/12 mandatory A/B/C/D commands passed with zero deselections, skips, xfails, or xpasses
• canonical final validator: ok=true, errors=[], promotion required and satisfied
• git diff --check passed; worktree clean

Impact

• benchmark implementation and measured values: unchanged
• training/backend behavior: unchanged
• test stability: removes an invalid cross-sample timing ordering while preserving the actual benchmark record contract
• support boundary: unchanged; the only Recommended combination remains ppo_torch + g1_walk_flat + mjwarp + device_resident

Integration

This child PR intentionally targets feat/issue-705-manager-mjwarp. Per #705, final PR #847 remains open for real main-targeted CI and user review; this PR does not approve or merge #847.