#46022 install: use one errno predicate so the macOS fallbacks fire for ENOTSUP

Before and after (darwin arm64 and x64, same results on both)

Before is build 128736: main plus tests only, macOS 26.4 arm64 and macOS 14 x64. After is build 129314: commit 23328ad8ee, macOS 15 arm64 and macOS 14 x64.

Cache and project on one HFS+ or exFAT disk image (hdiutil), one tarball dependency:

With the cache on APFS and the project on the image, both linkers exit 0 on main: clonefileat reports EXDEV first, and that check works. --backend=copyfile is the workaround for 1.4.0 to 1.4.3.

test/cli/install/bun-install-backends.test.ts on the darwin lanes: 31 pass and 4 fail on main (the 4 tests of the first push). 37 pass, 0 fail and 4 skip on this branch (the skips are the Windows tests).

clonefileat calls of one install, counted with the interposer on APFS (build 129314):

• 1 package: 1 refused call, then the fallback.
• 20 packages, hoisted: 3 refused (arm64) and 7 (x64), then the copyfile backend for the rest. Each thread that starts before the first refusal is stored is refused once.
• 20 packages, isolated: 3 refused (arm64) and 11 (x64), then the hardlink backend.
• Nothing refused: 20 successful clones for 20 packages.

The runs on disk images used a temporary probe test. It is not in the branch any more (removed in 2728dababa).

Errno of each call (0 is success):

When the cache is on the project's volume

• The user puts it there: BUN_INSTALL_CACHE_DIR, --cache-dir, install.cache.dir, or BUN_INSTALL, XDG_CACHE_HOME or HOME on that volume.
• Bun puts it there: [install.cache] disable = true or a cache directory that cannot be opened gives node_modules/.cache. No HOME, XDG_CACHE_HOME or BUN_INSTALL gives node_modules/.bun-cache (src/install/PackageManager/PackageManagerDirectories.rs).
• With install.globalStore the isolated linker clones into <cache>/links, so the volume of the cache alone decides.

A project on an external drive with the cache in the home directory never failed.

The five gates

Where the regression comes from: before the port bun.sys.E was Zig's std.posix.E, whose Darwin enum has OPNOTSUPP = 45 and nothing at 102. The port (#30412) made E the same type as SystemErrno, where EOPNOTSUPP = 102, and carried the arms over by name. It corrected the manifest save in src/install/npm.rs by hand in the same commit.

Left alone on purpose

• renameat_concurrently_without_fallback (src/sys/lib.rs) still names EOPNOTSUPP. It is inert on macOS: RENAME_EXCL answers EEXIST for a taken name before the file system is asked, so ENOTSUP means the target is missing, and RENAME_SWAP then answers ENOENT. The delete-then-rename path runs in both cases.
• Bun.write has its own fcopyfile gate that falls back on EBADF only (src/runtime/webcore/blob/copy_file.rs). #36692 has that one.
• The hardlink gates (map_linkat_err, needs_copy): #45761 has them.

Landing order

#45572, #45761 and #45929 each rewrite a line that this PR changes and type EOPNOTSUPP again there. Their hunks apply on main and not on this branch. With the last commit of this PR, such a line no longer builds for macOS. The resolved lines:

• Isolated clonefile arm with #45572: errno if bun_errno::is_not_supported(errno) => self.set_backend(InstallMethod::Hardlink),
• .bin gate with #45761: its hunk adds | sys::Errno::ENOSYS inside the matches!. && !bun_errno::is_not_supported(errno) stays.
• link_refusal in #45761: _ if bun_errno::is_not_supported(errno) => ...
• fcopyfile gate with #45929: Err(err) if bun_errno::is_not_supported(err.get_errno()) => {}

Open question: hard link or copy in the isolated linker

1.2.19 to 1.3.14 went from a refused clone to the hardlink backend in the isolated linker (#21122) and to a copy in the hoisted linker. This PR restores both. If a copy is wanted for both linkers, the isolated arm becomes InstallMethod::Copyfile, the isolated test expects nlink 1, and exFAT then works without #45761. The docs name only the copy fallback. I left them as they are until this is answered. If the hard link stays, docs/pm/isolated-installs.mdx, docs/pm/cli/install.mdx and docs/pm/global-cache.mdx each need one sentence.

The last commit

#[deprecated] on Darwin's EOPNOTSUPP = 102 makes rustc reject a bare use in code compiled for macOS (the workspace denies warnings). bun run rust:check-all shows it from Linux. Five places keep the name with an allow: the errno table, the predicate, the two socket gates that want 102 (open(2) in src/dotenv/env_loader.rs, listen(2) in src/runtime/socket/Listener.rs) and the rename gate. The manifest save and rename_no_replace already compared both values and now call the predicate, because the guard requires it. Checked by putting E::EOPNOTSUPP back at the fcopyfile gate: the darwin check fails with the note, the Linux check passes.

Measurements

CI artifacts are on a host that the machine that made this PR cannot reach, and it has no macOS SDK. So these are release objects of single crates (cargo rustc --release --emit=obj with the release flags, no LTO), main against this branch, compared with llvm-size and a per-function llvm-objdump diff.

• darwin arm64, bun_install __TEXT,__text: 2,327,924 to 2,328,052 bytes (+128). Task::run (isolated) +22 instructions of 3,331, install_into +5, install_with_clonefile_each_dir +2, link_bin_or_create_shim +1, rename_no_replace -3.
• darwin arm64, bun_sys __TEXT,__text: 71,076 bytes on both. copy_file_with_state +5 instructions.
• Linux x64: bun_install text 3,029,269 to 3,029,255 bytes (-14, all in rename_no_replace). bun_sys is identical.
• is_not_supported leaves no symbol: it is inlined.
• From each of the 5 clonefileat calls to its success branch the instructions are the same. So is fcopyfile.
• The .bin gate's errno check takes 9 instructions for EEXIST on darwin arm64, 7 on main. Linux is identical.
• Allocations added: 0.

Checks

• cargo check --workspace for aarch64-apple-darwin, x86_64-apple-darwin, x86_64-unknown-linux-gnu, x86_64-pc-windows-msvc and x86_64-unknown-freebsd, after the merge of main at 963452075e. cargo clippy for the touched crates on Linux and darwin.
• On the Linux debug build: bun-install-backends, bun-install-hardlink-fallback, dt-unknown-readdir, bun-link, bun-install-staging, isolated-relink, bun-pm-cache-pack, test/cli/run/env.test.ts.
• The interposer (test/fixtures/refuse-syscalls-shim.c) looks the names up before it refuses, as the kernel does. Its logic ran on Linux in a unit harness.

Prior work

The predicate and the guard arms come from commit 579f46cbdb3 on the branch robobun/0d97fe3b/install-link-or-copy-per-file, which was never opened as a PR.