AURA-KSP final-checkpoint CPU timing: revision supplement

The script, protocol and cohort are exact copies of the pre-measurement kit.
The recorded CPU run is supplied under received/ in this revision package.
Reaggregating those measurements needs no skeleton data, PyTorch or new inference.
See reproduction/REPRODUCE.md and analysis/cpu/audit_cpu_results.py.

To make a new measurement (distinct from reproducing the archived statistics):
1. Use Python 3.11 x64 in an isolated environment. Install:
   python -m pip install numpy==2.3.5 PyYAML==6.0.3
   python -m pip install torch==2.11.0 --index-url https://download.pytorch.org/whl/cpu
2. Copy the four published final checkpoints from
   research-assets/final/run/checkpoints/STREAM/final.pt
   into cpu_timing/models/STREAM/final.pt, for STREAM = joint, bone,
   joint_motion, bone_motion. The script checks all checkpoint SHA-256 values.
   Source: https://doi.org/10.5281/zenodo.22858090, release v1.0.1.
3. Extract the fixed official-training inputs from an authorized local prepared
   archive (the data are not distributed here):
   python extract_cpu_inputs.py --source /path/to/ntu120_xset_processed_v023.tar.gz --output selected_inputs
4. Software validation (synthetic inputs; no timing result):
   python cpu_benchmark.py smoke --output-root software_check
5. Measurement:
   python cpu_benchmark.py run --data-dir selected_inputs --output-root results

Run the commands from this cpu_timing directory. A new run uses local hardware
and therefore should not be expected to reproduce the recorded latency numbers.
Retain all valid runs, interruptions and reasons for any repeat. The fixed
protocol does not select a favorable run or discard timing outliers.

No training, quality evaluation or XSet test forward is performed. Both systems
use the same retained checkpoints, with fresh forwards in each timed call.
This is single-host, one-thread, batch-one resident-input CPU timing; it does not
remeasure final-model RTX 4090 latency, memory, energy or a complete video pipeline.
Models and NTU skeleton arrays are referenced rather than redistributed here.
Project and third-party notices accompany the unchanged runtime source.
