Handle all-NaN channels in calc_adc_params (fixes #485)

[Leonard013]

Fixes #485.

Why

Writing a signal whose channel is entirely NaN crashes:

import numpy as np, wfdb
wfdb.wrsamp('xxx', fs=500, units=['mV'], sig_name=['I'],
            p_signal=np.array([[np.nan]]), fmt=['16'])
# ValueError: cannot convert float NaN to integer

Two root causes in wfdb/io/_signal.py, both noted in #485:

1. The all-NaN guard never runs. calc_adc_gain_baseline has
   if pmin == np.nan: which is always False (NaN == NaN is False). For an
   all-NaN channel pmin/pmax are NaN, so execution falls through to the
   "regular varied signal" branch, computes a NaN baseline, and dies at
   baseline = int(np.floor(baseline)) with ValueError: cannot convert float NaN to integer.
2. Unsuppressed warning. calc_adc_params calls np.nanmin/np.nanmax on
   the all-NaN column, emitting RuntimeWarning: All-NaN slice encountered. The
   existing # Should suppress warning message. comment shows this was already
   intended to be silenced.

What

• calc_adc_gain_baseline: if pmin == np.nan: → if np.isnan(pmin):, so the
  pre-existing all-NaN branch (which sets adc_gain = 1, baseline = 1) actually
  executes.
• calc_adc_params: wrap the np.nanmin/np.nanmax calls in
  warnings.catch_warnings() + simplefilter("ignore", RuntimeWarning), scoped to
  just those two lines.
• Added import warnings.

Note on suppression method

np.errstate(...) does not silence this warning — "All-NaN slice encountered"
is raised through Python's warnings machinery (a RuntimeWarning), not the
floating-point error state. warnings.catch_warnings() is required. Verified
empirically.

Why the sentinel gain/baseline is correct (NaN → invalid-value round-trip)

WFDB stores a missing/invalid physical sample as the format's reserved digital
sentinel INVALID_SAMPLE_VALUE[fmt] (e.g. -2**15 = -32768 for fmt 16), the
lowest digital code — which is exactly why calc_adc_gain_baseline does
dmin = dmin + 1 (reserve the bottom code for NaN).

In adc (adc_inplace), NaN positions are recorded up front and, after the
gain/baseline multiply-add, overwritten with the sentinel
(np.copyto(p_signal, d_nan, where=nanlocs)). For an all-NaN channel every sample
is a nanloc, so gain/baseline are never reflected in the stored data — they only
need to be a finite gain and an in-range int baseline so the .hea is well-formed.
On read, dac maps d_signal == d_nan back to np.nan before applying
gain/baseline. Hence the round-trip is exact regardless of the sentinel
gain/baseline, and this is identical to how a partial-NaN channel's individual
NaN samples are already handled (test_physical_conversion: p_signal[5:10]=nan ↔
d_signal[5:10]=-32768).

Observed round-trip (fmt 16): all-NaN channel → header adc_gain=1.0, baseline=1,
every digital sample -32768, read back as all-NaN. Mixed 2-channel signal: normal
channel round-trips to max abs err 0.0; all-NaN channel reads back all-NaN. No
warnings.

Tests

Added to tests/test_record.py::TestSignal, following the existing
test_adc_gain_{max,min}_boundary write/read round-trip style:

• test_all_nan_single_channel — single all-NaN channel; asserts no
  RuntimeWarning on write and read-back is all-NaN.
• test_all_nan_mixed_channels — 2 channels, one all-NaN + one normal; asserts no
  RuntimeWarning, normal channel round-trips within quantization tolerance, NaN
  channel reads back all-NaN.

Both fail before the fix (ValueError: cannot convert float NaN to integer)
and pass after. Full tests/test_record.py, tests/test_multi_record.py,
tests/test_processing.py pass. black --check clean.

Notes for reviewers

• adc_gain=1, baseline=1 defaults are the pre-existing all-NaN branch values
  (the fix just makes the branch reachable). Any finite gain / in-range integer
  baseline round-trips correctly since the samples are all the reserved sentinel;
  happy to switch to baseline = 0 if preferred.
• Scope: the same np.nanmin/np.nanmax warning could arise in the
  e_p_signal (expanded) branch of calc_adc_params. The crash is already fixed
  there (that branch shares calc_adc_gain_baseline); warning suppression was scoped
  to the documented p_signal path (the one with the # Should suppress comment,
  exercised by Handle all-NaN channels in calc_adc_params #485). Can extend to the expanded path if wanted.
• docs/changes.rst is batched by maintainers at release time (current 4.3.1 has
  no unreleased section), so no changelog entry is included.

---

Prepared with AI assistance (Claude Code): the reproduction, root-cause analysis,
fix, and tests were AI-drafted; the reasoning and verification (including the
NaN→sentinel round-trip trace) are documented above.