[shinaoka]

Release t4a-faer 0.26.0 with t4a-faer-traits 0.24.2. The maintainer approved both versions on 2026-10-29.

• Why 0.26.0 rather than 0.25.3: threadteam is the only parallel route: drop Par::Rayon, the spindle path and the Rayon cost model #97 removed Par::Rayon, Par::rayon, ParShape::Rayon and ParShape::rayon after faer-v0.25.2. The README says published versions follow semver, and consumers such as tenferro-rs require ^0.25.0.
• t4a-faer-traits 0.24.2: faer-traits/ only gained rust-version = "1.89.0".
• Pins: threadteam 0.1.3 and t4a-private-gemm-x86 0.1.27 are the latest on crates.io.

Changes since 0.25.2 (from CHANGELOG.md)

Breaking: threadteam is the only parallel route (#97)

Par::Rayon(NonZeroUsize), Par::rayon(nthreads), ParShape::Rayon(NonZeroUsize) and
ParShape::rayon(nthreads) are removed, together with the spindle path and the Rayon cost model.
Par::Seq and Par::Team remain. To migrate:

• a one-off parallel call: faer::team::with_team(n, |par| { /* faer calls with par */ }) replaces
  Par::rayon(n);
• repeated parallel work: build a threadteam::Pool and Team once and lend Par::Team(..) to each
  call, par.reborrow() for successive calls (see "parallelism" in README.md);
• a workspace query: pass par.shape() (or ParShape::Seq) where ParShape::rayon(n) was used.

A build without the threadteam feature has Par::Seq only.

New

• faer::linalg::matmul::matmul_with_conj_uninit and MatUninitMut: a matrix product into
  MaybeUninit<T> storage, returned as an initialized MatMut; prior contents are never read
  (matmul: a MaybeUninit-safe overwrite destination (upstream faer-rs#269) #136, requested upstream as [Feature] Add a MaybeUninit-safe overwrite destination for matmul sarah-quinones/faer-rs#269).

Correctness

• Nonsymmetric EVD: the input is scaled into range as ?GEEV does (LAPACK validation: the nonsymmetric EVD fixtures exceed the GET22-style residual threshold (4 dtypes) #112); the eigenvector back
  substitution clamps relative to the eigenvalue, so residuals no longer grow with n (LAPACK validation: the nonsymmetric EVD fixtures exceed the GET22-style residual threshold (4 dtypes) #112); the real
  Schur deflation-window sort always ends (real Schur: the deflation-window sort never terminates when a 2 x 2 swap fails #144); the 2 x 2 block solves of the real back
  substitution follow ?LALN2 (real EVD: eigenvectors of strongly graded matrices have GET22 residuals up to 2.6e9 #145).
• SVD: the rotation path finishes when dqds does not converge (svd: DCHKBD type 16 returns NoConvergence at n = 127-129 since the dqds path (#52) #110); a retained Ritz pair of the
  partial SVD carries its residual estimate through the restart (svd: a retained Ritz pair carries its residual estimate through the restart (#102) #108).
• LU: the full-pivoting search breaks ties by column then row, independent of the SIMD width (lu: pivot-search ties by column then row, independent of the SIMD lane count (#102) #105, full-pivoting LU: ties are broken by SIMD lane, not by position (pivot choice depends on SIMD width) #106).
• Products with a diagonal read the depth in either direction (LAPACK validation: pivoted LDL leaves invalid metadata on the prefilled-subdiagonal fixtures (4 dtypes) #113).
• Sparse LU reports an exact zero pivot as LuError::NumericSingular in both backends (sparse LU: a numerically singular matrix factors as Ok, and the solve returns NaN (upstream #339) #132).
• Krylov solvers: the final residual is reported on NoConvergence, and BiCGSTAB stops at the last
  finite iterate after a breakdown (matrix_free: NoConvergence reports the initial residual, not the final one (upstream #328) #133, bicgstab: a breakdown (r̃ᵀv = 0) is not detected, and x becomes NaN (upstream #327) #134).

Parallelism and performance

• Every region runs at the widest width that pays, on a sub-team of that width (並列ゲートの dispatch cost が幅 6 で頭打ち、通れば常に全幅: p=32〜64 で floor 付近の region が過剰に開く #91); the team
  dispatch cost is read from threadteam (cost: read the team dispatch cost from threadteam 0.1.2; drop team_section_cost (threadteam-rs#6 stage 4) #104).
• No heap allocation inside a parallel region (par: no heap allocation inside per-step parallel sections (full_piv_lu best vector, for_each_child item/slot vectors) #111).
• Layout copies split along the destination's major dimension, first touch in parallel (linalg: split the layout copies along the destination's major dimension, first touch in parallel (#92 part 1) #99).
• SVD: blocked bidiagonalization, dqds shift and a padded work copy (perf(svd): blocked bidiagonalisation, dqds shift, padded work copy, caller-held worker group; values-only SVD at or below the vendor #52).

Dependencies and toolchain

• threadteam 0.1.3 (team route: some processes run every repetition at the 1-thread time (16T, broccoli) #107 wait fix) and t4a-private-gemm-x86 0.1.27 (LAPACK validation: pivoted LDL leaves invalid metadata on the prefilled-subdiagonal fixtures (4 dtypes) #113).
• MSRV is Rust 1.89 (rust-version = "1.89.0" in both crates).
• AVX-512 (x86-v4) no longer needs the nightly compiler but stays opt-in: enable the x86-v4
  feature (nightly is kept as its old name).
• t4a-faer-traits 0.24.2 only adds rust-version.

Verification

• cargo publish --dry-run -p t4a-faer-traits -p t4a-faer packages and verifies both crates (cargo 1.99.1).
• The code is unchanged from main (b903fa1), whose CI is green.

After merge

• Tag the merge commit faer-v0.26.0.
• From a clean detached worktree at the tag, publish t4a-faer-traits 0.24.2, then t4a-faer 0.26.0.
• Check the registry versions and dependencies.
• The downstream requirement updates (tenferro / tlinalg / tcontract) are handled separately.

🤖 Generated with Claude Code