Summary

Fixes #29 — ITD's itd_baseline_extract method crashed with a ValueError when processing signals where the number of local maxima and minima differed after boundary padding.

Root Cause

Lines 106-107 used np.append to flatten (index, value) pairs into 1D arrays:

LK1 = np.append(idx_min.reshape(-1, 1), LK1.reshape(-1, 1))  # flat 1D, length = 2*len(idx_min)
LK2 = np.append(idx_max.reshape(-1, 1), LK2.reshape(-1, 1))  # flat 1D, length = 2*len(idx_max)
LK = np.vstack((LK1, LK2)).T  # CRASH: requires equal lengths

After the boundary condition adjustments (lines 74-88), idx_min and idx_max could have different lengths (e.g., 14 vs 12 as reported), making np.vstack fail with:

ValueError: all the input array dimensions except for the concatenation axis must match exactly,
but along dimension 1, the array at index 0 has size 14 and the array at index 1 has size 12

Changes

pysdkit/_itd/itd.py:

• Use np.column_stack to build proper (position, value) 2D arrays for LK1 and LK2 control points, then np.vstack to merge them — works regardless of differing array lengths
• Fix boundary padding to use the actual signal value at the mirrored position (x[idx]) instead of reusing the first/last value from the wrong envelope
• Add guard for signals with no extrema (pure trends), returning the signal as baseline with zero rotation component
• Handle division-by-zero in piecewise slope computation when consecutive extrema have equal signal values
• Use np.concatenate with explicit int casting for index arrays to prevent float indexing errors
• Remove debug print statements

pysdkit/tests/test_itd.py (new):

• Add comprehensive test suite with 7 test cases:
  • Built-in test signals (3 cases)
  • __call__ interface
  • Signal reconstruction accuracy verification
  • Near-monotonic signal (the crash case from issue Issue with ITD method #29)
  • Pure trend signal (no extrema)
  • Asymmetric extrema counts (different number of maxima vs minima)
  • String representation

Test Plan

• All 9 new ITD tests pass
• All 25 existing EMD/CEEMDAN tests still pass (regression check)
• Reconstruction error < 1e-10 for all signal types
• Near-monotonic signals no longer crash
• Pure trends handled gracefully
• Signals with unequal maxima/minima counts work correctly