sigproc: fix OOB block read in decon_2D that crashes restore_baseline (#491)

[HaiwangYu]

Fixes #491.

Symptom

The ICARUS 2D signal-processing chain crashes with

Waveform.cxx: Throw in function percentile
Dynamic exception type: boost::wrapexcept<WireCell::ValueError>
[WireCell::tag_errmsg*] = empty waveform

thrown from OmnibusSigProc::restore_baseline via
OmnibusSigProc::decon_2D_tighterROI, intermittently ("event 2" in CI, not seen
in production).

Root cause

Every 2D-decon helper extracts the physical wires with

m_r_data[plane] = tm_r_data.block(m_pad_nwires[plane], 0, m_nwires[plane], m_nticks);

which assumes tm_r_data still carries the FFT wire-padding (m_fft_nwires
rows). For the ICARUS "twofaced" layout the induction planes are split into ≥2
sub-planes, so decon_2D_init() calls unpad_data(), which shrinks the array
to m_nwires rows before the final m_c_data = fwd_r2c(m_r_data). Starting the
block at row m_pad_nwires then reads m_pad_nwires rows past the end.

Observed at runtime (plane 0, TPC WW):

tm_r_data = (2112, 4296)   block(start_row=10, nwires=2112, nticks=4096)   fft_nwires=2132

→ rows [10, 2122) from a 2112-row array → 10 rows out of bounds.

Production libraries are built with -DNDEBUG, so the Eigen bounds check is
compiled out and the read silently returns adjacent heap memory, which is
occasionally NaN/Inf depending on allocator/run/data. That NaN reaches
restore_baseline(): median() → NaN, the |x-baseline|<500 cut rejects every
sample, and median() is then called on an empty vector → "empty waveform".

Fix

1. Root cause — derive the wire-axis padding offset from the actual row
   count at all decon_2D extraction sites (7 tm_r_data slices + the rawdecon
   finalization slice):

   const int roi_row_pad = (tm_r_data.rows() - m_nwires[plane]) / 2;
   m_r_data[plane] = tm_r_data.block(roi_row_pad, 0, m_nwires[plane], m_nticks);

   This equals m_pad_nwires when the FFT padding is present (unchanged
   behavior) and 0 once unpad_data() has removed it (in-bounds, correct
   wires).

2. Defense-in-depth — harden restore_baseline() against non-finite
   samples: drop NaN/Inf (zeroing in place) before the baseline estimate and
   guard the second median() against an empty selection. No behavior change
   for finite data.

Validation

Rebuilt libWireCellSigProc.so from the v0.36.1 source (the version the
failing icaruscode v10_20_09p01 loads) with Eigen assertions ON:

• Unpatched: aborts at the out-of-bounds .block() in
  decon_2D_tighterROI during the WW-TPC deconvolution on the CI input file.
• Patched: the same job runs both events through 2D signal processing with
  no assertion and no "empty waveform" (exit status 0), and the
  restore_baseline non-finite guard never fires — i.e. removing the overrun
  removes the NaN at its source.

The changed restore_baseline / decon_2D_tighterROI bodies are byte-identical
between this master branch and the validated v0.36.1 tree.

🤖 Generated with Claude Code

https://claude.ai/code/session_016Md5zsMdMEEgatTrumsBoV