PATCH NOTES (post-delivery, applied 2026-08-14)

Base: tahoe_axis_benchmark as delivered from the original compute session
(library modules at git commit 918fb12).

One change has been made to the delivered tree:

  tests/test_harness.py::test_basal_pc_features_recovers_rank
    The assertion was rewritten to compare Gram matrices (pairwise inner
    products of the PC scores) instead of reconstructing the centered matrix
    against loadings recomputed outside the function.

    Reason: basal_pc_features selects variable genes with np.argpartition,
    whose output ordering is not guaranteed -- numpy's introselect takes
    different code paths on different CPU architectures. The old assertion
    mixed PC scores computed in the permuted gene space with loadings
    recomputed in the original order, so it passed on x86_64 and failed on
    arm64. The model output is provably unaffected: a column permutation
    leaves the Gram matrix unchanged, hence the PC scores are identical up to
    sign, and all reported benchmark numbers are architecture-independent.

    Verified: 27/27 tests pass on both x86_64 (original session) and arm64
    (this verification), where the suite previously gave 26/27.

No library module, script, config, or lock file was modified. No result was
recomputed.
