Ignore class_values entries absent from the input array

[wrignj08]

Problem

clean_array raised KeyError when class_values named a class that does not occur in the input array — but only when smooth_edge_size > 0:

a = np.full((64, 64), 254, np.uint8); a[10:50, 10:50] = 1   # only class 1 present
clean_array(a, class_values=[0, 1], smooth_edge_size=3, ...)  # KeyError: 0
clean_array(a, class_values=[0, 1], smooth_edge_size=0, ...)  # fine

The smooth path in smooth_edges_to_codes did an unguarded value_to_code[cv_], and value_to_code is built solely from np.unique(array). The two sibling call sites already guarded this — multiclean.py filters target_codes on membership, and the background loop uses .get with a None check — so this was a missed case, not intentional strictness.

This mainly affects tiled processing, where one fixed class list is reused across tiles whose contents vary. The failure depends on the data, so it surfaces intermittently.

Fix

Filter target_class_values against the values present in the array once at the entry point. That covers all three call sites and skips the wasted cv2.morphologyEx call for a class whose mask is entirely zeros. background_class_values is derived as set(all) - set(target), so it is unaffected.

Tests

Added parametrized sweeps over dtype × smooth_edge_size × min_island_size × class_values. The existing point tests missed this because the bug lived in one corner of that grid, so the new tests assert invariants across it:

• absent classes never raise and never change the result
• if every requested class is absent, the array is returned untouched
• shape/dtype preserved and output values are always a subset of input values

Verified by reverting the fix: 121 of the new cases fail without it. Suite is 322 passed in ~1.7s.

🤖 Generated with Claude Code