[rwgk]

Summary

Document the maintenance policy for the staged rebuild of #2737: maintain both bindings majors on main, with explicit release branches for stabilization and urgent fixes. Related: #1199.

• Document independent release scope, major-line versus frozen patch maintenance, and reconciliation back to main.
• Replace automatic historical-branch backports with an explicit manual workflow that opens a reviewable PR for one existing release branch.
• Keep 12.9.x frozen except for emergency fixes using compatible historical tooling, and require reconciliation into the main-tree maintenance root.
• Record cross-root review and generator-provenance expectations. Mark deliberately duplicated test setup and require agents to compare every copy during development and review.

The full stack preserves current main's Lychee work, shared build backend, Python matrix, and platform improvements. The source import also includes the later 12.9.10 release and graph-callback fix absent from the frozen #2737 implementation.

Stack / Review Boundary

Layer 7 of 8 in native stack #3179. The following #3182 isolates the optional runtime maintenance fixes and contains the final integrated source state. The stack is rebased onto main dfad7e3658a9, preserving its Pathfinder release integration and NVML fixes. Earlier integrated validation remains recorded below; full hosted CI has not been rerun after this rebase.

Layer	PR	Review scope
1	#3172	Exact, inactive CUDA 12 source import
2	#3173	Source, build-helper, and test integration
3	#3174	Package selection, SCM, and developer environments
4	#3175	Build planning, wheels, sdists, and artifacts
5	#3176	GPU tests, source checks, coverage, and docs
6	#3177	Tagged-source release and artifact validation
7	#3178	Maintenance policy and release-branch workflow
8	#3182	CUDA 12 runtime backports and their regression tests

Suggested Review Order

1. .github/RELEASE-bindings.md
2. ci/README.md
3. .github/workflows/backport.yml
4. AGENTS.md

Validation

Current local validation after rebasing onto main: full pre-commit passes on #3173 and the final #3182 head; 422 CI-tool tests and 55 subtests pass; six resolver/checker CLI integration cases pass. Six layers preserve their pre-rebase diffs. The release layer reconciles main's Pathfinder flow with tagged-source routing, with an independent review finding no merge regressions. The CUDA 12 source tree is unchanged, and main's three updated NVML files are preserved exactly.

The earlier integrated QA results below precede this rebase. Full hosted CI has not been rerun.

• Full pre-commit run --all-files --show-diff-on-failure passed at the initial integrated head; focused pre-commit checks passed for all seven follow-up files.
• 402 CI-tool tests and 55 subtests; 279 build-hook tests (9 skipped).
• CUDA 12.9 QA: 1,615 Pathfinder tests, 410 bindings tests in each normal/PTDS stream mode, 3,835 Core tests, and both Cython suites (9 bindings tests and 1 Core test).
• CUDA 13.4 QA: 1,615 Pathfinder tests, 669 bindings tests in each normal/PTDS stream mode, 4,002 Core tests, and both Cython suites (9 bindings tests and 1 Core test).
• Independent beta/RC and final metapackage rehearsals: four sdists and eight wheels, including rebuilds outside Git without version overrides. Both bindings metadata backends preserve release-tag README source links.
• Follow-up checks: 360 parallel CUDA 12 capture-test invocations on CPython 3.12 (2/4 threads, normal/PTDS), six marker-classification cases, and 12 CI label/matrix tests.

The initial hosted run passed all 28 wheel builds, sdists and documentation checks, and most GPU tests. Three capture failures and two infrastructure findings were addressed and checked; that run is superseded.

At the last observation, run 39549648508 for commit fb5ccedcb64c44bd171d8bfa9169b2a5c88bde65 recorded 102 successful jobs. Every CUDA 12 row fully passed, including all free-threaded Python targets.

Full hosted CI remains incomplete. The queued RTX 2080 job is being ignored at maintainer request. A separate Windows Core/CUDA 13 Python 3.14t job, using published bindings 13.4.1, failed with an access violation in test_incremental_ltoir_round_trip; a targeted retry request was rejected by GitHub (HTTP 403: the workflow is still running), so one retry remains blocked until the workflow completes. Matching main and #3165 jobs passed the same test with the same bindings and NVRTC/NVVM/nvJitLink versions. The root cause remains unclassified. This PR remains draft.

Checklist

• Relevant existing and focused regression tests cover the integration behavior.
• This description identifies the layer's scope and validation boundary.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[rwgk]

/ok to test 9cf3778

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-3178/
https://nvidia.github.io/cuda-python/pr-preview/pr-3178/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-3178/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-3178/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[rwgk]

/ok to test fb5cced