CI: slow Checkout cuda-python on Windows — submodule clone not partial-cloned #3351

In run 41718385081, Checkout cuda-python ranges from ~30s to 3m42s across jobs, dominated by the cccl submodule clone.

Fetching submodules sub-group, same cccl commit ba8eb88, same command:

• win-64 py3.11: 132s (job 126207494611)
• win-64 py3.12: 14s (job 126207494463)
• win-arm64 py3.11: 129s (job 126207493919)
• linux-64 py3.12: 73s
• linux-aarch64 py3.12: 88s

Platform isn't the dominant factor — identical windows-2022 runners spanned 14s–132s on the same commit. The variance is in the unfiltered submodule clone against GitHub's pack server.

Root cause: actions/checkout's filter input is applied only to the top-level fetch, not to submodule updates. The submodule call is a plain git -c protocol.version=2 submodule update --init --force (git-command-manager.ts:455-465). Upstream issue: actions/checkout#1714 (open).

Proposed fix — do the submodule step manually so it gets --filter=blob:none --depth=1:

- uses: actions/checkout@v5
  with:
    submodules: false
    fetch-depth: 1
- run: git submodule update --init --depth=1 --filter=blob:none --recursive

This also caps the long tail on Linux jobs.

Alternative, lower-blast-radius: keep submodules: true and add fetch-depth: 1 — --depth does propagate through submodule update (git-command-manager.ts:459), so it caps most of the cost without changing the action surface.

-- Leo's bot