[MRG] Fix sparse simplex projection with oversized budgets - #876

Types of changes

Bug fix: use ordinary simplex projection for an unconstrained sparsity budget, focused regressions, and a release note.

Motivation and context / Related issue

Fixes #875. The documented behavior for max_nz greater than the projected dimension is ordinary simplex projection. Instead, the row slice only contains the actual width while the threshold index array has length max_nz, producing a broadcasting error. When the budget is at least the row width, use the existing ordinary simplex projection directly; axis=0 and axis=None inherit this through their existing recursion. This avoids indexed mutation on TensorFlow tensors. Smaller budgets retain the sparse projection path.

How has this been tested

• Built POT's native extensions from this source checkout using MSVC 2022, OpenMP, Cython and the repository's Eigen submodule on Windows/Python 3.12.
• On unmodified master 98d09a1, the added regression yields 3 failures/3 controls with NumPy. Restoring that exact baseline function in memory in a second environment yields 6 failures/6 controls across NumPy and PyTorch CPU. No source file is altered by that baseline harness.
• After the fix, the complete test/test_utils.py module passes: 120 tests in the NumPy environment; 182 tests (2 warnings) in the NumPy/PyTorch 2.10 CPU environment. Expected values are analytic simplex solutions, covering all three axis modes with budgets equal to or larger than the projected dimension.
• Configured pre-commit hooks applicable to all three changed files pass; git diff --check passes. Hooks reporting no matching files are not counted as executed checks.
• Focused Sphinx autodoc/Napoleon build of the changed public API passes with warnings treated as errors. The harness references the existing algorithm footnote; its first isolated build warned that the existing footnote was unreferenced. The repository docstring is unchanged.
• Not run: full package suite, full project documentation/gallery, GPU, or the TensorFlow/JAX/CuPy backends. Zero/negative sparsity-budget validation is outside this change.

PR checklist

• I have read the CONTRIBUTING document.
• The documentation is up-to-date with the changes I made (check build artifacts). The documented contract already states this behavior; only the focused API build was run locally.
• All tests passed, and additional code has been covered with new tests. The complete utility module and regression passed; the full package suite was not run.
• I have added the Issue fix to RELEASES.md.

AI assistance disclosure

This contribution was investigated, implemented and validated autonomously with Codex assistance on behalf of AHMETHAKANBEZIR1. It has not received independent human code review. The commit records Codex as a co-author. Maintainer review is requested; local test evidence is not a claim of upstream CI success.

Current base validation

Merged master c02a25c, retaining both release-note entries and upstream device-allocation fixes. Complete utility module: 124 passed with NumPy; 190 passed with NumPy/PyTorch 2.9.1 CPU. Focused sparse-simplex tests in the NumPy/TensorFlow environment: 21 passed. Applicable configured hooks and diff checks pass. This update did not run the full package suite or GPU tests; CI is pending.

20 October master integration

Merged current master 5cb82c9, preserving both the existing fix release note and the upstream API-documentation release note. The only conflict was in RELEASES.md; production code and regression tests merged without conflict. Complete utility module with NumPy: 124 passed. Focused sparse-simplex tests with NumPy/TensorFlow: 21 passed. Applicable configured pre-commit hooks and diff checks pass. These are Windows/Python 3.12 CPU checks; no new full-package, GPU, or PyTorch run is claimed for this maintenance update. GitHub CI is pending. Autonomous Codex contribution disclosure above still applies.

20 October CI completion check

Current-head CircleCI documentation build and documentation artifact checks now pass. The labeler also passes. No remote unit-test result is present in the reported contexts; the local numerical test results above remain the validation evidence. The pull request remains open pending maintainer review.