convert: honor dither argument with ADAPTIVE palette (#5836)

Tests/test_image_convert.py

@@ -377,3 +377,28 @@ def test_matrix_identity() -> None:
     # Test list
     converted_im = im.convert("RGB", list(identity_matrix))
     assert_image_equal(converted_im, im)
+
+
+def test_adaptive_convert_honors_dither(tmp_path: Path) -> None:
+    # Regression test for #5836: convert("P", palette=ADAPTIVE, colors=N,
+    # dither=...) must apply dithering, not just nearest-color mapping.
+    im = hopper("RGB")
+
+    no_dither = im.convert(
+        "P", palette=Image.Palette.ADAPTIVE, colors=8, dither=Image.Dither.NONE
+    )
+    with_dither = im.convert(
+        "P",
+        palette=Image.Palette.ADAPTIVE,
+        colors=8,
+        dither=Image.Dither.FLOYDSTEINBERG,
+    )
+
+    # Both should produce valid palette images with the same palette size.
+    assert no_dither.mode == "P"
+    assert with_dither.mode == "P"
+    assert len(no_dither.getpalette()) == len(with_dither.getpalette()) == 24
+
+    # Dithered and non-dithered results must differ. Histograms should spread
+    # differently because dithering distributes error across neighboring pixels.
+    assert no_dither.histogram() != with_dither.histogram()

src/PIL/Image.py

@@ -1204,7 +1204,13 @@ def convert_transparency(
                     raise ValueError(msg)
 
         if mode == "P" and palette == Palette.ADAPTIVE:
-            im = self.im.quantize(colors)
+            # quantize() maps each pixel to the nearest palette entry without
+            # dithering, so it silently drops the dither argument. Build the
+            # adaptive palette first, then re-convert with dither so the
+            # dither argument is honored (see #5836).
+            adaptive = self.im.quantize(colors)
+            palette_im = self._new(adaptive)
+            im = self.im.convert(mode, dither, palette_im.im)
             new_im = self._new(im)
             from . import ImagePalette