Fix lda.default "group means are numerically identical" error on macOS/Windows

[Copilot]

Newer MASS binaries (macOS/Windows) apply a strict rank check in lda.default that throws "group means are numerically identical" when the between-class scatter matrix has rank 0 in the sphered space. This is triggered by features that are linearly dependent in the within-group-centred space — including the original unit test data where the within-group deviations of feature1 and feature2 are identical, causing the group means to project to the same point after within-class sphering.

Changes

• R/lefser.R — filterCollinearFeatures: New internal function that detects and removes linearly dependent features before LDA. It computes the within-group-centred feature matrix, runs QR decomposition to identify redundant columns, removes them, and emits a named warning so users know which features were affected and can take action (e.g. using get_terminal_nodes):

Warning: Linearly dependent features removed before LDA: feature2.
Consider using `get_terminal_nodes` to reduce feature redundancy.

• R/lefser.R — ldaFunction: Calls filterCollinearFeatures() on the feature matrix before fitting the LDA model, replacing the earlier tryCatch+jitter approach. Also removes the now-redundant commented-out createUniqueValues call in lefser().

• tests/testthat/test-lefser.R: Updated the score test to use non-degenerate data (5 samples per class with distinct within-group variation patterns for each feature), so no filtering occurs and LDA runs normally. Added a dedicated test "ldaFunction warns and removes within-group collinear features" that verifies the collinearity warning is emitted for the original pathological anti-symmetric data.

[lwaldron]

@copilot I don't like the idea of masking an error (identical group means) by adding jitter. The error occurs because of true collinearity, and adding jitter only masks that collinearity. A correct fix would be to filter (features?) to actually address the collinearity. The user should be aware of and have control over such filtering.

[Copilot]

@copilot I don't like the idea of masking an error (identical group means) by adding jitter. The error occurs because of true collinearity, and adding jitter only masks that collinearity. A correct fi...

Replaced the jitter approach with a proper filterCollinearFeatures() function (commit 1cc7e6b).

How it works: Before calling lda(), the function computes the within-group-centred feature matrix and runs a QR decomposition to detect linearly dependent columns. Redundant features are removed and a warning is emitted listing the removed feature names along with a suggestion to use get_terminal_nodes:

Warning: Linearly dependent features removed before LDA: feature2.
Consider using `get_terminal_nodes` to reduce feature redundancy.

This surfaces the collinearity to the user rather than silently masking it. The tests were updated accordingly: the score test now uses non-degenerate data (different within-group variation patterns per feature), and a dedicated test verifies the warning is emitted for genuinely collinear inputs.

[lwaldron]

Not a fan of these "fixes" - the real problem was artificial rank-deficient test data (each class is rank 1), relying on a seemingly previously undefined behavior of MASS::lda in such cases but which should have errored. Apparently this is a new feature of MASS to correctly throw this error. I will replace the test data with something non rank-deficient offline.
test_that("ldaFunction correctly identifies classes and calculates scores", {
class_A_data <- data.frame(
feature1 = c(10, 11, 12),
feature2 = c(1, 2, 3), # -> c(2, 0, 4)
class = "A"
)

class_B_data <- data.frame(
    feature1 = c(1, 2, 3), # -> c(2, 0, 4)
    feature2 = c(10, 11, 12),
    class = "B"
)