EDPW v2_0_01 -- INDEPENDENT NUMERICAL AUDIT (DEVELOPMENT)
2026-10-08

Independent Python/NumPy reference check:
  45 Direct-EDPW / chirp-convolution Fast-EDPW comparisons: PASS
  30 Goertzel bin cases versus NumPy FFT: PASS
  6 CZT contiguous-band cases versus NumPy FFT: PASS
  w=N recovers the full DFT: PASS for tested power-of-two sizes
  Maximum relative Direct/Fast discrepancy: 7.908e-13

This verifies independently transcribed formulas, NOT execution of the
MATLAB source itself. MATLAB R2016a and GNU Octave were not available.
The MATLAB public SELFTEST has been expanded to include w=1, w=N,
non-radix-2 N=300, selected-block equality, and exact selected DFT bins.

RELEASE GATES STILL OPEN:
  - Run SELFTEST and RUN_ALL in actual MATLAB R2016a.
  - Confirm figures/tables and all experiments against manuscript.
  - Independently assess spectral-detection statistics and fair task costs.
  - Account for FFT-kernel precomputation and measured runtime separately.

No claim of published/validated experimental superiority is warranted.

2026-10-08: MATLAB self-test includes direct operator-matrix construction and
normalized Frobenius identity across 7 small record sizes. This test is
SOURCE-LEVEL ONLY until run in MATLAB R2016a.
