Keep rank-deficient lm and glm fits predictable

[EmilHvitfeldt]

Ticks only the first box of #308. Items 2 (label parsing) and 3 (prediction intervals and case weights) are untouched and stay open.

The problem

parse_model_lm() called qr.solve(qr.R(model$qr)) on every parse. A rank-deficient fit has a singular R factor, so the parse aborted for a model whose fitted values are well defined, even though tidypredict_fit() never touches the QR decomposition.

The issue understates the reach: two everyday shapes fail, not one.

set.seed(1); n <- 120
d <- data.frame(x1 = rnorm(n), x2 = runif(n, 0, 10))
d$y <- 2 * d$x1 - 0.5 * d$x2 + rnorm(n, sd = 0.3)

tidypredict_fit(lm(y ~ x1 + xdup, transform(d, xdup = d$x1)))    # duplicated column
tidypredict_fit(lm(y ~ x1 + xconst, transform(d, xconst = 1)))   # zero-variance predictor
#> Error: Unable to calculate inverse of QR decomposition.

A predictor with no variance is not mentioned in the issue and is the more realistic way to hit this. Both fail identically for glm(), gaussian and binomial. predict() handles all of them.

The fix

Drop the NA coefficients R leaves for the terms it could not identify, as predict() does, and build the inverse from the leading rank rows and columns of the R factor: exactly the columns whose coefficients survive. That block is not singular, so prediction intervals keep working for rank-deficient models as well, which was better than expected.

The QR failure no longer aborts the parse. qr_inverse_lm() returns NULL when no inverse can be formed and te_interval_lm() reports it with a cli message, which is the only place the QR is required. Previously that path failed with Must supply `.init` when `.x` is empty. This also removes the grepl("singular matrix", cnd$message) test on an untranslated English message and the bare stop(cnd) re-raise, both noted as secondary in the issue.

Measured

Max absolute difference against the modelling package's own predict(), 120 rows:

model	shape	fit	interval
lm	duplicated column	8.9e-16	8.9e-16
lm	zero-variance predictor	8.9e-16	8.9e-16
lm	mpg ~ cyl_f * gear_f, 1 aliased coef	0	n/a
glm gaussian	duplicated column	8.9e-16	n/a
glm binomial	zero-variance predictor	1.1e-16	n/a
lm full rank (regression check)	mpg ~ wt + disp	3.6e-15	0

Intervals compared against predict(interval = "prediction")[, "upr"].

Other classes checked

parse_model_lm() is shared by lm, glm and rq. quantreg::rq() refuses to fit a singular design itself and stores no $qr, so it takes the same NULL path as before: unchanged. The parsnip linear_reg() lm and glm engines route through the same parser and now work on rank-deficient fits (4.4e-16 against predict()). The other callers of build_terms() (lda, qda, fda, sda, sparsediscrim, mixOmics, multinom, glmnet) pass qr = NULL and were not touched.

Tests

New tests in tests/testthat/test-model-lm.R for both shapes, for lm and glm, compared against stats::predict(). Each was confirmed to fail before the fix (4 failures on the reverted source). The old #124 snapshot test asserted the abort that this PR removes, so it now asserts the fitted values match predict() instead.

devtools::test(): FAIL 0 | WARN 13 | SKIP 0 | PASS 2002 (the warnings are the stale local H2O cluster). devtools::check() is clean apart from a hidden-directory NOTE from the worktree and one H2O REST failure in test-h2o.R:142, both unrelated to this change.