Changes in version 0.1.0                        

Initial public release.

Core architecture

  - S4 classes adgeMatrix (dense) and adgCMatrix (sparse) extend the
    Matrix package with backend-dispatch slots for transparent GPU
    acceleration.
  - Five backend targets: CPU (default), MLX (Apple Silicon), OpenCL
    (CLBlast), Metal (Apple sparse), and ArrayFire (portable GPU).
  - Three precision modes: "strict" (float64 CPU), "exact" (float64
    CPU-pinned), and "fast" (float32-permitted GPU).
  - Automatic dispatch planning with calibration thresholds, resident
    memory management, and predictable CPU fallback.

Dense linear algebra

  - matmul, crossprod, tcrossprod, covariance, dist_matrix —
    backend-dispatched with GPU-resident fast paths.
  - chol_factor, lu_factor, svd_factor — factorization caching with
    solve, project, and reconstruct methods.
  - eigh, am_svd, rsvd, block_lanczos — eigendecomposition and
    randomized/iterative spectral methods.
  - sinkhorn — doubly-stochastic scaling with GPU-resident key-chain
    loop.
  - ewise, am_sweep, segment_sum, segment_mean — element-wise and
    reduction operations.
  - woodbury_solve, woodbury_logdet — Woodbury matrix identity solvers.

Statistical models

  - many_lm — fit multiple linear models via GPU-accelerated QR with
    coefficient caching.
  - lm_fit, array_lm, wls_fit — single and weighted least squares.
  - ridge_path, ridge_fit — ridge regression over lambda grids.
  - lm_loo_cv — leave-one-out cross-validation via QR downdate.
  - pca_coef — PCA coefficient extraction from cached SVD factors.

Sparse operations

  - adgCMatrix with backend-dispatched SpMV and SpMM via OpenCL, Metal,
    and MLX resident paths.
  - block_lanczos and svd_factor subspace iteration on sparse inputs.

Backend packages (separate installation)

  - amatrix.mlx — Apple MLX backend via mlx-c; full dense + sparse
    support.
  - amatrix.opencl — OpenCL + CLBlast backend; dense BLAS,
    factorizations, and sparse products.
  - amatrix.metal — Apple Metal backend; sparse-first with MPSGraph
    kernels.
  - amatrix.arrayfire — ArrayFire backend; portable GPU with safety
    gates for Apple Silicon.

Benchmark harness

  - tools/benchmark-regression.R — canonical regression harness with
    per-backend worker isolation, direct-file and source-entry parity,
    and automatic baseline comparison.

Observability and honest defaults

  - amatrix_backend_health_probe() — run a canary op against a backend
    and record its health (healthy / unhealthy:<reason>). Every backend
    is probed on first use; subsequent routing decisions respect the
    probe.
  - amatrix_backend_status() now reports per-backend health and
    health_reason columns alongside capabilities, features, and
    precision modes.
  - amatrix_fallback_log() and amatrix_fallback_log_reset() — structured
    log of every dispatch fall-through from a backend to CPU. A
    non-empty log after a clean conformance run is a stop-ship
    condition.
  - amatrix_benchmark_report() — reads tools/baseline.csv and the cached
    calibration, returning cold-vs-warm timings, warm/cold ratios,
    calibrated thresholds, and gpu_wins flags per (op × backend).
  - Calibration cache is tagged with a sys_hash over OS, machine,
    platform, and R version. Stale caches from a different machine are
    rejected automatically instead of producing wrong routing.
  - Auto-fallback with telemetry: every backend error in
    amatrix_dispatch_op is re-signaled as a calling-style condition (so
    withCallingHandlers can observe the original error class) AND emits
    an amatrix_fallback condition with structured metadata, before
    routing to CPU.
  - mode = "balanced" is now deprecated. It was never fully implemented
    (routed to CPU under the hood) and now maps to "exact" with a
    one-time-per-session deprecation warning. Use "exact" or "fast".

Error handling and diagnostics

  - User-facing errors in wrapper functions now emit classed conditions
    (amatrix_bad_arg, amatrix_bad_backend, amatrix_backend_exists,
    amatrix_subspace_error) with call = NULL so messages don't leak
    internal call stacks. tryCatch(..., amatrix_bad_arg = ...) and
    expect_error(class = "amatrix_bad_arg") both work.

Quality machinery

  - planning_docs/quality-tracking.md is the authoritative quality doc.
    It carries the coverage matrix (every exported op × four test types
    + benchmark row + per-backend tier), the stop-ship rules, and the
    honest backend tier assessment.
  - tools/audit-dispatch.R — static AST-based audit that enumerates
    every required dispatch signature for %*% / crossprod / tcrossprod
    on mixed plain/amatrix pairs and fails the PR gate if any is
    missing.
  - Residency tripwire (options(amatrix.residency.tripwire = TRUE) or
    AMATRIX_RESIDENCY_TRIPWIRE=1) increments a counter on every real
    GPU-to-host copy; the conformance suite asserts zero trips after a
    clean run.
  - Three-lane CI: PR gate (.github/workflows/R-CMD-check.yaml +
    fast-gates job), nightly stress (nightly-stress.yaml), release gate
    (release-gate.yaml with required reviewers environment).