NaN values appear nearly black with fire on Matplotlib 3.11 #848

When plotting a 2D array with imshow and cmap="fire", NaN values appear nearly black instead of transparent.

I encountered this when replacing nonpositive values with NaN for logarithmic normalization.

Minimal example

import numpy as np
import ultraplot as uplt

data = np.array([
    [np.nan, np.nan, np.nan],
    [np.nan, 0.1, 1.0],
    [np.nan, 10.0, 100.0],
])

fig, ax = uplt.subplots()
im = ax.imshow(
    data,
    cmap="fire",
    norm="log",
    vmin=0.1,
    vmax=100.0,
)

print(im.cmap.get_bad())
# Approximately [0.0148, 0.0142, 0.0142, 1.0], rather than transparent.

fig.show()

Expected behavior

NaN values should be transparent by default.

Environment

• UltraPlot 2.7.2
• Matplotlib 3.11.2
• Python 3.14

Possible cause

PerceptualColormap._init() in UltraPlot 2.7.2 calls _set_extremes() before setting _isinit = True.

The relevant Matplotlib implementation changed between these versions:

• Matplotlib 3.10.9: _set_extremes() writes the special colors directly into the lookup table, without checking _isinit.
• Matplotlib 3.11.2: _set_extremes() calls _update_lut_extremes() only when self._isinit is true.
• The actual lookup-table writes are now in _update_lut_extremes().

Consequently, on initial construction with Matplotlib 3.11.2, UltraPlot's call to _set_extremes() does not populate the special-color entries.
Those entries retain the initial np.ones values, including alpha = 1, and are then converted from HSL to RGB, producing an opaque, nearly black color.

The colormap's _rgba_bad is still (0, 0, 0, 0), but get_bad() returns an opaque, nearly black color.
In a process-local comparison, replacing only Matplotlib's _set_extremes() implementation with the one from 3.10.9 restored transparent NaN colors.

Edited: corrected both source-code links and added the version comparison above. The original line anchors pointed to unrelated code.

Workaround

Setting the bad color after initialization restores transparency. This was verified through rendering:

im.cmap.get_bad()  # Ensure initialization.
im.cmap.set_bad("none")

Related issues

Related: #831, but this concerns NaN transparency rather than colormap lookup.
I searched existing open and closed issues and pull requests and did not find a matching report.