[BenjaminDEMAILLE]

Fixes satijalab/seurat#10445, reported with a clear diagnosis and a workaround; this is that workaround made part of subsetting.

The state

RunUMAP(return.model = TRUE) keeps the embedding it was fit on in the reduction's misc$model. Subsetting the reduction leaves it untouched:

ncol(subset)                                     50
nrow(Embeddings(subset[["umap"]]))               50
nrow(Misc(subset[["umap"]], "model")$embedding)  80

Why it matters

Projecting a query onto that reduction, which is what MapQuery() and ProjectUMAP() do, finds the query's neighbours among the reduction's current cells and then reads their coordinates out of the model. Every index past the subset therefore points at a different cell than the one that was matched. Nothing warns; the query simply lands in the wrong place, which is the distorted plot in that issue.

Measurement

Projecting a reference's own cells back through its own model, so each should land on its own coordinate, with a subset that is not a prefix of the original (a prefix hides the bug, since the first N rows still line up):

model	median distance to own coordinate	as a share of the embedding's spread
stale, 80 rows	3.184	13.7%
subset, 50 rows	0.579	2.5%

Change

subset.DimReduc() subsets the model's embedding to the cells the reduction keeps. Everything else in a uwot model describes features or parameters (n_neighbors, metric, scale_info, pca_models) and is left alone. A reduction with no model, or a model with no embedding, is untouched.

Tests

tests/testthat/test_subset_reduction_model.R: the model's embedding follows the reduction's cells and matches its coordinates, the rest of the model is unchanged, and a reduction with no model still subsets.

Two assertions fail without the change. SeuratObject suite: 183 passing, 0 failures. Seurat's own suite against this build: 1184 passing, the only 2 failures being ones that predate it and are unrelated (Read10X_Image, satijalab/seurat#10454).

🤖 Generated with Claude Code

[BenjaminDEMAILLE]

Superseded by #318, which groups the correctness PRs into one so that a single review covers the subject. The commit is carried over unchanged; closing here to keep the queue down rather than because anything was wrong with it.