[soodoku]

Closes #117.

Every HAL basis function switches on as a covariate increases, so negating a covariate changes the penalty and the fit (example in #117). This adds an opt-in reflect argument to fit_hal() and enumerate_basis().

What changes

• enumerate_basis(..., reflect = TRUE) builds the basis on cbind(x, -x). basis_of_degree() and enumerate_edge_basis() get an internal reflected_from argument, which drops tensor products that pair a covariate with its own reflection.
• fit_hal(..., reflect = TRUE) makes the design matrix from cbind(X, -X), names the reflected columns -x1, -x2, ... (so summary() reads, e.g., I(-x1 >= -0.5)), and stores reflect in the fit. predict() augments new_data the same way.
• reflect = TRUE with formula stops with an error.
• The default, reflect = FALSE, is unchanged.

The quantizer and the edge basis work column by column, so negating a covariate only swaps the two copies of it: the reflected basis is the same set of functions under any sign coding, with no change to quantizer().

Tests (tests/testthat/test-reflect.R)

• Fits along a fixed lambda path, zero and first order, are unchanged when two of three covariates are negated (max difference < 1e-3 at thresh = 1e-12; observed 2e-5 to 3e-5).
• The same comparison with the default basis changes by more than 1e-2, so the test can fail.
• No basis function pairs a covariate with its reflection, and every covariate enters in both directions.
• predict() on new data and summary() work; formula with reflect = TRUE errors.

Checks

• devtools::test(): 52 passing, 0 failures.
• devtools::check(build_args = "--no-build-vignettes"): 0 errors, 4 warnings, 1 note, identical to master on the same machine (vignettes not built, Rcpp build version, hal in tests, a lost brace in formula_helpers.Rd).
• styler leaves the changed lines alone. devtools::document() regenerated only the four affected Rd files; I left RoxygenNote at 7.3.1.

Targets master, since devel is contained in it; happy to retarget.

🤖 Generated with Claude Code