Hellings-Downs optimal statistic from NANOGrav 15-year products (Sandbox Physics, AC61)

Contents
  optimal-statistic/  six Python files, standard library only
  verification/       four Python files for the independent checks, standard library only
  frozen-inputs/      pulsar positions and noise-weighted inner products derived from
                      NANOGrav 15-year data set v2.1.0 (Zenodo 10.5281/zenodo.16051178)
                      and NG15 noise spectra v1.0.0 (Zenodo 10.5281/zenodo.8092346),
                      both CC-BY-4.0; no raw residual tables are included
  benchmarks/         the saved run: inputs.json, outputs.json, failure-cases.json and
                      independent-verification.json

Recompute the statistic (about 13 s on Python 3.9.6):
  cd pulsar-timing-optimal-statistic/optimal-statistic
  python3 reproduce_optimal_statistic.py ../frozen-inputs ../my-run

Rerun the independent checks (about 40 s):
  cd ../verification
  python3 run_independent_verification.py ../frozen-inputs/ng15-pulsar-positions.json ../my-run/independent-verification.json

To rebuild the frozen inputs, download both Zenodo records and run
  python3 extract_ng15_frozen_inputs.py DATASET_DIRECTORY NOISE_SPECTRA_FILE OUTPUT_DIRECTORY

Scope: a simplified optimal statistic with five gamma = 13/3 signal bins, epoch-averaged
residuals, a three-column spin timing model and three per-pulsar noise variants. It does
not reproduce the published amplitude; the recovered A^2 is about 1/12 of 5.76e-30, and
the cause is not verified. Quote significance against the phase-shift and sky-scramble
nulls, not the amplitude. The injection study is an offline simulation with synthetic
coefficients at the real pulsar positions.
