RadioDataset samples are commonly split I/Q with shape (2, T) (row 0 = I,
row 1 = Q). `sample_spectrogram_plot` assumed 1-D complex, so `len(sample)`
returned 2 (the I/Q axis) — the compatibility gate saw `2 < 32` and every
dataset showed "doesn't have sufficient signal data for spectrogram
visualization", and the fallback path raised "need at least 32 samples, got 2".
Normalize each sample to a 1-D complex signal before measuring/plotting:
- add `_to_complex_1d` (complex any-shape -> flat; (2, ...) I/Q rows and
(..., 2) I/Q cols -> I + jQ; real 1-D -> real signal; None on empty).
- compatibility gate and `sample_spectrogram_plot` normalize first, then use
the true length; plot returns the styled "Not Available" figure for
unusable/too-short (<32) samples.
- `_compute_spectrogram` measures length via reshape(-1) (belt-and-suspenders).
Tests: parametrized `_to_complex_1d` and `sample_spectrogram_plot` over
complex 1-D, (2, T) rows, (T, 2) cols, real 1-D, and (2, 4, 256) multi-channel
I/Q (all render), plus (2,)/empty (still "Not Available"). 16 tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
