scanpy pca silently subsets to var.highly_variable after gene injection
In scanpy >=1.10, sc.pp.pca/sc.tl.pca defaults mask_var to 'highly_variable' whenever that column exists in adata.var. If you subset adata to an externally supplied gene list (e.g. Seurat's VariableFeatures) or overwrite adata.X with an external scaled matrix, genes not flagged by scanpy's own HVG call are silently dropped from PCA. Pass mask_var=None (or reset var['highly_variable']=True) when injecting gene lists or matrices. Diagnose it by comparing to an exact numpy SVD on the full vs reduced gene set: a perfect match to the reduced set gives it away.
The mechanism, the fix and the diagnostic all reproduce, but the behaviour predates 1.10. Test: 500 cells x 2000 genes, normalize_total + log1p, highly_variable_genes(n_top_genes=300), then subset to an 'external' list of 150 scanpy HVGs + 350 non-HVGs, scale, pca(n_comps=10, arpack).
scanpy 1.12.5: top-3 variance 2.2648 2.1953 2.1625. Exact SVD on the 150 flagged genes gives 2.2648 2.1953 2.1625, and on all 500 genes 3.8641 3.7689 3.7470. Only 150 loading rows are nonzero. With mask_var=None the result matches the full set (3.8641 ...).
scanpy 1.9.8 (py3.10, numpy<2): identical numbers. The parameter there is use_highly_variable (its default None uses the column if it exists), and mask_var does not exist. So on 1.9.8 the fix is use_highly_variable=False. I only tested 1.9.8 and 1.12.5.