This looks good to me. One minor concern is that as recordings get bigger, both check_spike_times_without_nans and check_ascending_spike_times load the whole array into memory and will take some time to run.

Typical large recording today:

• ~1,000–5,000 units (multi-probe Neuropixels)
• mean rate ~5 Hz, session ~1–2 h (~3,600–7,200 s)
• → 5,000 × 5 × 7,200 ≈ 1.8e8 spikes

Some labs are doing many-hour recordings, so on the order of 1e9 spike times is plausible.

With 1e9 spike times:

data[:] materializes 8 GB (float64) and np.isnan allocates another ~1 GB bool array, so ~9 GB transient. On a 16 GB laptop running the inspector that risks an OOM.

On my Mac M1 laptop, it would take about 35 seconds to read gzip-compressed data[:] + check, with a warm cache.

I think that's OK, but something to keep in mind as data gets bigger.

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.87%. Comparing base (13d0d67) to head (4e6efde).
⚠️ Report is 6 commits behind head on dev.

@@            Coverage Diff             @@
##              dev     #689      +/-   ##
==========================================
+ Coverage   79.44%   82.87%   +3.43%     
==========================================
  Files          48       48              
  Lines        1897     1904       +7     
==========================================
+ Hits         1507     1578      +71     
+ Misses        390      326      -64