feat: lightweight PCA backend (scipy/numpy SVD)

[adRn-s]

Replace the sklearn PCA and StandardScaler in Correlation.plot_pca with a scipy/numpy SVD implementation, reproducing sklearn's output (column standardization with population std, deterministic svd_flip sign convention, explained_variance_ from singular values). Drop the now-unused pandas and scikit-learn dependencies.

[adRn-s]

Added PCA test coverage + fixed three inherited plot_pca bugs (a04e88c)

While building a regression test suite for plotPCA (to guard the sklearn → scipy/SVD switch), we discovered that the SVD rewrite produces the same default output as sklearn — good — but that three pre-existing bugs in the surrounding plot_pca logic were carried over unchanged. This commit adds the tests and fixes all three.

Discoveries

• The scipy/SVD PCA is numerically equivalent to sklearn for the default path (matches to rtol=1e-5), differing only by per-principal-component sign — expected gauge freedom. Confirmed by running the same suite against both implementations.
• --transpose was fully broken (in both sklearn and scipy versions): Wt = np.dot(m, Wt.T) raises a shape mismatch whenever features ≠ samples, i.e. essentially always. This is the option meant to "match what R does," and it could never run.
• --log2 and --rowCenter were silent no-ops. They mutated self.matrix after m had already been copied during --ntop filtering, so the transforms never reached the PCA. Output was byte-identical to default.
• --ntop below the sample count crashed the scatter with an IndexError (fewer components than samples to lay out).

Fixes

1. Transpose — dropped the broken np.dot; U * S already gives each sample's projection onto the PCs, so we just orient it as (components, samples).
2. log2 / rowCenter — now applied to a float copy of the matrix, before variance filtering, so both flags actually affect the result (and self.matrix is no longer mutated as a side effect).
3. Small --ntop — added guards that sys.exit with a clear message instead of crashing.

Tests (test_plotPCA.py, 14 passing)

Sign-invariant coordinate + eigenvalue regressions (default and transpose), --ntop behavior, variance/eigenvalue consistency, --PCs selection, and CLI validation exits. The sign-invariance helper normalizes along the correct component axis (columns for the untransposed table, rows for the transposed one) and stores raw goldens, so the suite is a fair cross-implementation net — it passes against both the sklearn and scipy backends rather than only agreeing with itself.

Note

pca_back currently fails to build its Rust extension locally (maturin errors), so the PCA tests were run via the working env pointed at this branch's pure-Python source. The rest of the suite (anything needing the compiled deeptools.hp) was not exercised here.

[WardDeb]

ty, looks good to me !