[alphaville]

Main Changes

Target version: v2.0.0

Added

• scripts/run_python_tests.sh runs the Python pytest suite with configurable interpreter/build paths and optional C++ exchange tests.
• File-I/O tests cover shapes, precision, binary layout, non-contiguous inputs, malformed files, filesystem errors, and actual C++/Python exchanges using temporary fixtures.
• LeastSquaresBatchedWorkspace<T> preallocates reusable status storage and one-entry matrix-pointer arrays only for batches of one.
• Svd::rankRelative(rtol, atol = 0) adds GPU rank calculation using per-matrix thresholds max(atol, rtol * sigmaMax) and validated tolerances.
• Build/test and CUDA memory-check scripts run C++/Python tests, repeated leak checks, and save sanitizer logs.
• Size-safe Session::recordAllocation(size_t), recordDeallocation(size_t), and counted cudaRelease methods.
• DTensor::getRowsInto(destination, rowsFrom, rowsTo, matIdx) validates and copies rows into preallocated storage on the source stream.
• DTensor::leastSquaresBatched(b, infoArray) supports reusable device status storage without per-call allocation.
• Session::numStreams() exposes the initialized stream count.
• Configurable GPU benchmarks compare batched GEMM and row-extraction methods, with correctness checks, host/GPU timings, and CSV results.
• Regression tests cover validation, files, GEMM coefficients, stream ordering, ownership, rank tolerances, and reusable buffers.

Changed

• Breaking: replace incrementAllocatedBytes(int) with separate size-safe allocation/deallocation methods.
• Breaking: copy assignment deep-copies; move assignment transfers storage and empties the source. Use slices for aliases and deviceCopyTo for preallocated copies.
• Reshape retains and reuses owned matrix-pointer capacity where possible.
• Column-major uploads use the supplied host vector directly; only row-major uploads allocate a conversion buffer.
• setStreamIdx returns a reference for lvalues and moves rvalues, avoiding implicit allocation and deep copying.
• getRows uses one asynchronous 2D copy and inherits the source stream; getRowsInto preserves the destination's stream index.
• Tensor constructors validate dimensions, allocation overflow, vector length, and storage mode before allocating.
• Validate reshape sizes, slices, random bounds, row extraction, debug access, empty factoriser inputs, and QR output shapes.
• Reject zero streams and reconfiguration after initialization; validate stream indices against the initialized count.
• Document numerical preconditions, buffer lifetimes, stream dependencies, reusable workspaces, and binary-format assumptions.
• Ignore generated build files, dependencies, binaries, and temporary test artifacts, including in-source builds.

Fixed

• Nullspace projection uses constructor-allocated scratch and stream-ordered copy-back to avoid GEMM aliasing; tests cover dense batches, repeated calls, streams, and allocation reuse.
• Givens annihilation uses the identity rotation for an already-zero pair, avoiding NaNs and preserving both rows; covered for float and double.
• Python file I/O rejects incomplete headers, payload-size mismatches, zero dimensions, empty/scalar arrays, and object/zero-size dtypes before loading or overwriting data.
• Python binary loading preserves column-major element order for single matrices, with a rectangular round-trip regression test.
• Batched Cholesky and least-squares support single matrices using constructor-allocated pointer storage without allocations during reusable solves.
• Size QR workspace for the largest requirement across factorisation, least-squares, and Q extraction; add large square/tall regressions for both precisions.
• Share stream configuration across translation units using an inline static Session member, with a two-file regression test.
• Prevent SVD constructor exception leaks using a unique_ptr owner and raw observer; verify cleanup with allocation-failure tests.
• Tensor constructors release GPU storage on subsequent host-allocation failures; fault-injection tests verify cleanup for float, double, and int.
• Session shutdown completes pending work, releases library handles, and destroys owned CUDA streams.
• Single and batched Cholesky queries, factorisations, solves, and status tensors use the matrix-selected stream.
• Host transfers, scalar reads, initialization, and pointer uploads complete on their selected stream; transpose results inherit the source stream.
• Device copies wait for prior destination work and complete on the source stream before returning.
• QR debugging-helper uploads complete on the selected solver stream before returning, fixing intermittent stream-ordering failures.
• Track owned bytes and pointer capacity independently of shape, count successful allocations/frees, and preserve accounting across reshape, views, and moves.
• Reset Svd::rank() counters on each call; SVD factorisation, rank kernels, and results use the input stream.
• Assignment releases previous storage and safely handles metadata, self-assignment, and empty tensors; deep copies complete on the source stream.
• Enable testing at the project root so ctest --test-dir build discovers the test suite.
• Make the Python-format binary-reader test independent of generated Python files and the working directory.
• Reject malformed text and incomplete or incorrectly sized binary files; report write failures and handle short filenames and empty data safely.
• Avoid null-pointer arithmetic when slicing a single-matrix tensor along the matrix axis.
• QR operations use the matrix-selected cuSolver/cuBLAS stream; getQR synchronizes before reading R on the host.
• Givens coefficient kernels use the selected cuBLAS stream, check launch errors, and restore the previous pointer mode.
• Make test-stream setup independent of execution order and add nonblocking-stream regressions for QR, Givens, and row extraction.
• Correct the row-slicing test to select a column before slicing its rows.

Associated Issues

• None for now