Release 0.7.1¶

SOLVAX 0.7.1 adds the generated-block foundation needed by memory-bounded kinetic solvers:

  • generated truncated block elimination shares one LU solve between Schur and multiple right-hand-side updates;

  • block_tridiag_matvec and block_tridiag_relative_residual provide an independent full block-row diagnostic;

  • vector, multiple-RHS, jit, vmap, reverse-mode, float32, float64, low-order, factor-reuse, and transpose contracts are covered by tests;

  • a generated kinetic example and versioned CPU/GPU benchmark artifacts are included;

  • the unused Lineax dependency and the corresponding operator-interface overclaim are removed;

  • repository tests explicitly import src/solvax, preventing an older global installation from satisfying CI.

On the checked-in production-shaped benchmark, CPU warm runtime changes by +0.36%, within the no-regression gate. An RTX A4000 multi-RHS workload is 5.53% faster, while cold compilation increases from 0.53 s to 0.65 s. These are hardware-specific measurements, not universal performance claims.