plotPCA: write and plot the per-sample loadings in the untransposed layout

[cindykrafft]

Fixes #1459.

• Does the PR contain new feature? No
• Does the PR contain bugfix? Yes — fixes plotPCA (4.0.0 re-implementation): the default table and plot hold the PC scores of the first bins, not the loadings of the samples #1459
• Does the PR contain documentation changes? A CHANGES.txt bullet only
• Does the PR contain changes to the galaxy wrapper? No

The re-implemented Correlation.plot_pca set Wt = U * S for both layouts. With the samples as the variables (the default layout) U * S holds the PC scores of the rows of the matrix — one row per selected bin — so the "Component i" row of --outFileNameData and the plotted points were the first rows' scores across the components, labelled with the sample names (row norms 1.5–1.9 instead of 1; the groups that PC1 separates in the loadings are separated by no row of the table). The help documents "the loadings for each sample in each principal component"; those are the rows of V^T, which 3.5.x wrote. --transpose (each sample's projection, (U * S).T) was already right and is unchanged, as are the eigenvalues, the variance fractions and the scree plot. The "not enough principal components to plot N samples" exit is removed: every component row has one loading per sample, so --ntop below the sample count plots fine as long as the requested PCs exist (the remaining check). Review notes: the golden test_plotPCA_default.png/.tsv and test_plotPCA_ggplot.png/.tsv showed the bin scores and are regenerated (the corrected plots were 6.8 and 5.7 RMS from the old goldens, under the image tolerance of 50, which is why the image tests did not catch this); test_plotPCA_ntop_smaller_than_samples now expects one loading per sample (a 2 × 8 table for --ntop 2, plot included) and test_plotPCA_requested_PC_beyond_ntop_errors_cleanly keeps the clean-exit coverage (--ntop 2 --PCs 1 3).

Test: test_plotPCA_default_table_holds_sample_loadings — the rows of the default table are orthonormal (W @ W.T == I) and PC1 equals a numpy SVD's V^T[0] on the same standardised rows, sign-invariantly. On master it fails (36/36 elements of W @ W.T mismatch) together with the adjusted --ntop 2 test (SystemExit: Not enough principal components); with the patch test_plotPCA.py + test_plotcorrelation.py 20 passed, correlation.py doctests pass, full pytest suite passes (numbers in test-runs.txt).

Found in Mytochondria, a volunteer project that checks the numerical core of research software and verifies every finding by execution (methods and harnesses: https://github.com/cindykrafft/mytochondria/tree/main/audits/deeptools)

---

Generated by Claude Code

---