• Closes Per-point size array is cumulatively reordered across multi-color subplots #4024
• Tests included or not required because: the fix is a one-line variable rename (loop-local _size instead of mutating size); existing tests pass and the bug is visual (per-point size arrays scrambled across subplots).
• Release notes not necessary because: this is a minimal internal bugfix with no API change.

When sc.pl.embedding (or sc.pl.umap, etc.) is called with multiple color keys and a per-point size array, the size variable is cumulatively reordered across loop iterations:

# line 299 in scatterplots.py
if isinstance(size, np.ndarray):
    size = np.array(size)[order]  # mutates loop variable!

Each subplot computes its own order (for z-ordering by color value), but applies it to the already-reordered size from the previous iteration. This means:

• Subplot 1: correct sizes
• Subplot 2: sizes scrambled by order1 × order2
• Subplot 3: sizes scrambled by order1 × order2 × order3
• etc.

Fix: use a loop-local _size variable so each iteration reorders from the original size array. The same pattern is applied to the scatter call and add_outline code that consume the size.

Minimal reproduction:

import scanpy as sc
import numpy as np

adata = sc.datasets.pbmc3k_processed()
sizes = np.random.default_rng(0).uniform(10, 200, size=adata.n_obs)
sc.pl.umap(adata, color=["louvain", "n_genes", "n_counts"], size=sizes)
# Dot sizes visibly differ across panels despite being the same array