[ucgmsim]

Closes #236

What was wrong

IM/snr_calculation.py:calculate_snr had two related bugs:

1. The "noise duration must be at least 1s" check compared noise_duration
   (a sample count, equal to tp) directly against 1, instead of
   noise_duration * dt against 1.0 seconds. A noise window of e.g. 0.5s
   (tp=50, dt=0.01) was silently accepted; the check only ever fired when
   tp <= 0.
2. tp was never validated against the waveform length nt. If
   tp >= nt, the signal slice became empty and the call failed deep inside
   ims.fourier_amplitude_spectra with an unrelated OverflowError
   (cannot convert float infinity to integer) instead of a clear error at
   the point of the bad input.

Also fixed the accompanying docstrings: SNRResult.signal_duration /
noise_duration were documented as seconds but are actually raw sample
counts (now typed and documented as such, per the issue's suggestion — the
SNR calculation only uses their ratio, so the unit choice doesn't affect
results), and the tp parameter was documented as float but is used as an
integer index (now documented as int).

Fix

• Added an upfront if not (0 < tp < nt): raise ValueError(...) check in
  calculate_snr.
• Changed the noise-duration check to noise_duration * dt < 1.0, matching
  the documented "less than 1s" rule.
• Updated SNRResult and calculate_snr docstrings accordingly.

Test

Added tests/test_ims.py::test_snr_noise_duration_in_seconds (a 1000-sample,
dt=0.01 waveform with tp=50, i.e. 0.5s of noise, now raises ValueError
instead of being silently accepted) and
tests/test_ims.py::test_snr_tp_validated_against_record_length (parametrized
over tp of 0, -1, 1000, 1200 against a 1000-sample waveform, all of
which now raise a clear ValueError instead of tp >= nt crashing with
OverflowError downstream). Both tests failed before the fix for the reasons
described in the issue, and pass after it.

For the reviewer

• tp == nt - 1 (e.g. tp=999 for nt=1000) is intentionally still
  accepted: it satisfies 0 < tp < nt and produces a very short (2-sample)
  signal window, per the issue's suggested fix of validating 0 < tp < nt
  rather than requiring a minimum signal length. Worth double-checking this
  matches the intended behaviour.
• signal_duration/noise_duration are kept as sample counts (not
  converted to seconds) to avoid changing the numeric values returned by
  calculate_snr beyond what the issue describes; only the type annotation
  and docstrings were corrected.

Cost $0.48, 27 turns.

🤖 Generated with Claude Code

[lispandfound]

@joelridden point 1 seems significant, no?