Fix psrflux loading bugs affecting the secondary-spectrum axes

[danielreardon]

Background

Found while investigating a user report of a "misaligned" secondary spectrum in high-frequency-resolution FAST data (PSR B1933+16, ~500 Hz and ~2.9 kHz channels).

I audited the whole chain from loading a psrflux dynamic spectrum through to plotting the secondary spectrum. calc_sspec and plot_sspec came out clean — verified empirically:

• zero-delay and zero-Doppler land on exactly the right bin for both even and odd nf/nt (the transform is zero-padded to a power of two, so input parity is irrelevant);
• pre-whitening introduces no shift — convolve2d(..., mode='valid') moves the sample grid by half a pixel, but taking |F|² discards that linear phase (synthetic features land on the identical bin with prewhite on and off);
• plot_sspec uses centres_to_edges + pcolormesh consistently with plot_dyn, so cells are centred on their coordinates, and the fitted-arc overlay shares that convention (its vertex sits exactly on the DC cell).

All of the real defects were in the loader.

Fixes

1. remove_short_subs deleted far more than it should (High)

The loop used <= together with a redundant, always-true sdt >= 0 guard. Once the surviving sub-integrations became uniformly spaced, sdt fell to 0 and the condition collapsed to dt0 - dt <= 0, so it kept deleting uniformly spaced subints.

Reproduced: a 6-subint file with one short leading subint lost 4 of them (only 2 survived), silently truncating the observation and its time axis. It now stops as soon as the first subint is no longer a short outlier.

2. Channel bandwidth off-by-one (Medium)

psrflux frequencies are channel centres, so the spacing is span/(nchan-1), but df was computed as span/nchan. df is now taken from the median difference between channel centres (also robust to an occasional missing channel).

For the bundled 400 MHz / 512-channel example file this changes df from 0.77972 to 0.78125 MHz (= 400/512, exactly). Since the delay axis is built as tdel ~ 1/df, this directly mis-scaled the secondary-spectrum delay axis — by ~1/(nchan−1), so significant for small channel counts and negligible for very large ones.

3. Non-uniform channels were silently assumed uniform (High, when it applies)

The ACF and secondary-spectrum FFTs assume a uniform frequency grid, but nothing checked. Loading now warns when the channel spacing is non-uniform (gaps or stitched sub-bands), and separately when the number of unique channel centres doesn't match the file's channel count.

4. Channel-ordering detection made independent of bw/df

Descending-frequency files are now detected from the first channel block rather than inferred from a negative bandwidth, so the flux-row flip no longer depends on how bw/df are derived. Verified to agree with the previous behaviour on the bundled example file.

Verification

• New tests/test_dynspec_loading.py (12 tests): remove_short_subs removes exactly the short subint / leaves a uniform cadence intact, df matches the true spacing for several channel counts and for the real example file, non-uniform channels warn (and uniform ones don't), descending and ascending files keep each flux value matched to its (frequency, time), and the delay axis reflects the corrected df.
• Full suite: 86 passed (74 existing + 12 new) — no regressions.
• flake8 CI gate (E9,F63,F7,F82) returns 0; the new test file is style-clean.
• Checked against the bundled J0437-4715 file: nchan/nsub/dt/tobs/bw unchanged, no spurious warnings, frequencies correctly ascending.

Note for users

The df change slightly alters the delay-axis scaling of previously computed secondary spectra (it is now correct). The effect is ~1/(nchan−1) — noticeable for low channel counts, well under 0.01% for the very large channel counts typical of baseband data.

🤖 Generated with Claude Code

---

Generated by Claude Code

[danielreardon]

Fixes for independently identified dynspec loading bugs.