fix(cargo-oxide): scaffold and clone from NVIDIA/cuda-rust

[alejandro-soto-franco]

What this adds

cargo oxide new writes the repository address of cuda-device and cuda-host into every new project's manifest, and the backend fallback clone fetches the same address when no cuda-oxide dependency pins a commit. Both still named the pre-move NVlabs/cuda-oxide repository, which resolves only through GitHub's redirect and leaves every new lockfile recording the old source. Both now name NVIDIA/cuda-rust.

cargo oxide new demo

Before  cuda-device = { git = "https://github.com/NVlabs/cuda-oxide.git" }
        source = "git+https://github.com/NVlabs/cuda-oxide.git#bc46b1ff..."   (via redirect)
        checkout ~/.cargo/git/checkouts/cuda-oxide-<hash>/bc46b1f/

After   cuda-device = { git = "https://github.com/NVIDIA/cuda-rust.git" }
        source = "git+https://github.com/NVIDIA/cuda-rust.git#bc46b1ff..."    (direct)
        checkout ~/.cargo/git/checkouts/cuda-rust-<hash>/bc46b1f/

cargo oxide new demo && cd demo
cargo oxide build   # builds the backend from the checkout Cargo made for cuda-device

What we fixed

• GIT_REPO in scaffold.rs, interpolated into both scaffold templates, now reads https://github.com/NVIDIA/cuda-rust.git.
• The fallback git clone --depth 1 in backend.rs, used only for projects with no cuda-oxide dependency, clones the same address.
• The backend_source.rs module diagram, its test constant and the matching README diagram follow: the checkout is cuda-rust-<hash>, and the backend crate and the toolchain pin sit under cuda-oxide/ in it.
• A scaffold test asserts the emitted git = "https://github.com/NVIDIA/cuda-rust.git" lines in both templates and rejects the old address.

Implementation

• Backend discovery never reads the URL: dependency_source_from_metadata keys on the package name and manifest_path, git_source_rev keeps only the #<sha> fragment, and checkout_root walks up to the first directory holding Cargo.toml plus the codegen crate, bounded by .git or .cargo-ok.
• codegen_crate_in_checkout prefers cuda-oxide/crates/rustc-codegen-cuda and falls back to the historical flat crates/rustc-codegen-cuda; pinned_channel walks from the codegen crate upward, so the component nightly wins over the stable pin at the git root. The same two functions serve the fallback clone.
• The backend cache is keyed by commit (source-rev.txt), not by URL, so cached backends and existing projects are unaffected.
• The historical issue permalink to NVlabs/cuda-oxide/issues/96 in the README stays, per the repository guide.

Verification

Checked at 746d60eb:

• ✅ cargo test -p cargo-oxide --all-targets: 243 passed; 244 with the new URL test.
• ✅ cargo clippy -p cargo-oxide --all-targets -- -D warnings, cargo fmt --check, rustdoc with -D warnings, and all 18 scripts/check-*.sh guards pass.
• ✅ grep -rn NVlabs crates/cargo-oxide/src returns nothing; the remaining hits are the README install line (docs(cuda-oxide): link the book and crate READMEs to NVIDIA/cuda-rust #1491) and the Tracking: Tile-to-SIMT Interop (cuda-oxide as SIMT participant) #96 permalink.
• ✅ In an isolated CARGO_HOME: cargo oxide new writes the new address in both templates; cargo generate-lockfile records git+https://github.com/NVIDIA/cuda-rust.git#bc46b1ff for cuda-device, cuda-host and cuda-macros in a cuda-rust-* checkout; cargo oxide build builds the backend from that checkout's cuda-oxide/crates/rustc-codegen-cuda under nightly-2026-10-11 and caches it with source-rev.txt = bc46b1ff.
• ✅ The old-address baseline binary scaffolds NVlabs/cuda-oxide.git and resolves the same commit through the redirect; git ls-remote on both URLs returns the same HEAD.
• ✅ A shallow clone of NVIDIA/cuda-rust builds the backend offline from its nested codegen crate under the component nightly, which is the fallback route.
• ✅ Auto-merges with fix(cargo-oxide): take the scaffold's host crates from cuda-host's git source #1488, fix(cuda-oxide): point repository metadata and the ICE report URL at NVIDIA/cuda-rust #1490 and docs(cuda-oxide): link the book and crate READMEs to NVIDIA/cuda-rust #1491 in either order.

Caveats

• A scaffolded project still fails to compile at this head with two cuda-core crates in its graph; that is fix(cargo-oxide): take the scaffold's host crates from cuda-host's git source #1488's fix and is independent of the address.
• A graph that mixes the old and new address for the same crate resolves two checkouts and cargo oxide refuses it; [patch."https://github.com/NVlabs/cuda-oxide.git"] pointing the crate at the new address collapses the graph to one source.
• Machines that already hold the old-address git database fetch the repository once more under the new name; the backend cache is reused because it is keyed by commit.
• The remaining old-address mentions are the install lines (docs(cuda-oxide): link the book and crate READMEs to NVIDIA/cuda-rust #1491) and the manifest repository fields and license CSV (fix(cuda-oxide): point repository metadata and the ICE report URL at NVIDIA/cuda-rust #1490).

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[nihalpasham]

/ok to test 3280e67

[nihalpasham]

Checked at 3280e67c:

• Discovery never reads the URL, so cached backends and existing projects are unaffected; a fresh scaffold now records NVIDIA/cuda-rust in its manifest and lockfile.
• I added one commit on top: the two discovery diagrams still drew the old checkout directory and flat crate layout, and a test now asserts both templates name the new repository and never the old owner.
• 244 cargo-oxide tests, clippy, fmt, rustdoc and the guards pass. CI is green on this head.