Release v5.6.0 · michael-denyer/jamma

Added

jlinalg.dsyrk accepts a caller-owned output buffer. The signature is now dsyrk(X, *, out=None, beta=0.0), computing out = X @ X.T + beta * out. The existing one-argument call is unchanged. Batch, streaming, and LOCO kinship accumulation use dsyrk(X, out=K, beta=1.0) in place of K += dsyrk(X), which removes the n_samples x n_samples temporary that the old pattern allocated on every batch.

Mode 4 computes its REML and MLE coarse-grid brackets in one pass. The two likelihoods share the same three SNP-varying reductions at each grid lambda and differ only in their tail, so the one-covariate split/SoA kernel now derives one canonical Pab per grid point and selects both brackets from it. Golden-section refinement and final evaluation remain independent per likelihood. Mode-4 output is bitwise unchanged. The general-covariate kernel is not affected.
jlinalg ABI version is 13. The bump is internal to the C extension; ABI mismatches trigger the existing automatic recompile.

dsyrk no longer reads its output buffer when beta is zero. BLAS defines beta == 0 as "C is not read", precisely so the caller may pass uninitialized memory. For a zero-width X the native path scaled the freshly allocated output instead of zeroing it, and because NaN * 0.0 is NaN, a reused allocation leaked its previous contents into the result. It is zeroed again.
dsyrk rejects an unaligned out instead of silently ignoring it. An unaligned but otherwise valid buffer was converted to an aligned copy, so the result was written to the copy and the caller's array left untouched. dsyrk(X, out=K) is K now holds by construction.

No public Python API breaks. dsyrk(X) still works and the new arguments are keyword-only.

Full changelog: v5.5.0...v5.6.0