Treat the WASM/WASI build target as cfg(unix) so filesystem tools need no per-tool mode-bit patches

uutils/coreutils' src/uucore/src/lib/features/fs.rs から始め、registry/native/patches にある参照先の WASI パッチを which.rs と builtins のエントリポイントと比較します。ここで説明されている target-spec、build-std、preview2 の代替案を調査します。互換性と保守に関する影響を含め、1 つの方向性を選択して文書化すれば完了です。

Summary

We build coreutils (ls, stat, chmod, …) for wasm32-wasip1. Because that
target is not cfg(unix), every crate's #[cfg(unix)] filesystem code —
the code that reads real st_mode permission bits — is compiled out, and the
#[cfg(not(unix))] fallback (which fabricates permissions from a single
readonly boolean) is compiled in. To get native-Linux-accurate output we've
had to hand-patch each tool to bypass that fallback and fetch mode bits from the
sidecar via a host_fs.path_mode import (landed in #268 / #269).

This issue tracks the larger question: should we make the WASM build target
present as cfg(unix) (a Linux-like target) so these tools "just work" with no
per-tool patches, instead of maintaining a growing pile of #[cfg(target_os = "wasi")] shims?

Why a non-Linux target breaks filesystem tools

uucore has two implementations of display_permissions, cfg-gated:

#[cfg(unix)]                       // reads REAL bits
pub fn display_permissions(md, ...) -> String {
    display_permissions_unix(md.mode() as mode_t, ...)   // st_mode
}

#[cfg(not(unix))]                  // WASI lands HERE — fabricates
pub fn display_permissions(md, display_file_type) -> String {
    let write = if md.permissions().readonly() { '-' } else { 'w' };
    format!("{file_type}r{write}xr{write}xr{write}x")     // r & x HARDCODED on
}

Upstream root cause (uutils/coreutils, uucore 0.5.0):
https://github.com/uutils/coreutils/blob/main/src/uucore/src/lib/features/fs.rs

Consequences of being cfg(not(unix)):

• md.mode() isn't even callable — std::os::unix::fs::MetadataExt doesn't
  exist in the WASI std.
• Rust's own std::os::wasi::fs::MetadataExt deliberately exposes
  dev/ino/nlink/size/atim/… but no mode(), because WASI preview1's
  filestat struct has no mode field. So the bits are absent at the std layer
  regardless of cfg.
• Net effect: ls -l printed -rwxrwxrwx (or -r-xr-xr-x if readonly) for
  every file — a uniform string derived from one boolean.

Note this is degradation, not deliberate stripping: the tools have "WASI
support," it's just a lossy fallback nobody wired to a host that has the bits.

Evidence: the per-tool workarounds this forces (all at 8145b07)

Each of these exists only because the target isn't cfg(unix):

• ls — inject a host_fs.path_mode import + mode_for_path and swap
  display_permissions → display_permissions_unix:
  https://github.com/rivet-dev/secure-exec/blob/8145b07ebcd1f6553a417a009091c53d7610a04f/registry/native/patches/crates/uu_ls/0001-wasi-host-fs-mode-display.patch#L16-L74
• stat:
  https://github.com/rivet-dev/secure-exec/blob/8145b07ebcd1f6553a417a009091c53d7610a04f/registry/native/patches/crates/uu_stat/0001-wasi-metadata-compat.patch
• chmod:
  https://github.com/rivet-dev/secure-exec/blob/8145b07ebcd1f6553a417a009091c53d7610a04f/registry/native/patches/crates/uu_chmod/0001-wasi-compat.patch
• First-party shims hit the same wall — which and the shell builtins each
  carry their own #[cfg(target_os = "wasi")] host_fs extern:
  https://github.com/rivet-dev/secure-exec/blob/8145b07ebcd1f6553a417a009091c53d7610a04f/registry/native/crates/libs/shims/src/which.rs#L16-L58
  https://github.com/rivet-dev/secure-exec/blob/8145b07ebcd1f6553a417a009091c53d7610a04f/registry/native/crates/libs/builtins/src/lib.rs#L14-L27
• And the C side already has the mirror-image fix in wasi-libc (which Rust can't
  reach, because Rust std bypasses libc stat() on wasip1):
  https://github.com/rivet-dev/secure-exec/blob/8145b07ebcd1f6553a417a009091c53d7610a04f/registry/native/patches/wasi-libc/0016-host-fs-mode-and-chmod.patch

Every new fs-touching Rust tool we add will need another such patch.

Options

A. Force cfg(unix) on the current target (cheap, does NOT work)

Passing --cfg unix via RUSTFLAGS flips the cfg on our crates, but the
#[cfg(unix)] path does use std::os::unix::fs::MetadataExt, which the
precompiled WASI std doesn't contain → unresolved import. And even if it
resolved, the WASI Metadata's underlying filestat has no st_mode to
return. Rejected.

B. Custom Linux-like target spec + -Z build-std (the real "treat it like native")

Define a target derived from wasm32-wasip1 with target-family = ["unix", …]
and recompile std/core so Rust's fs routes through the libc-backed unix fs
backend (calls libc stat(), reads st_mode). Then the wasi-libc mode patch
(0016) flows up into Rust automatically and all #[cfg(unix)] tool code works —
no per-tool patches. Blast radius, however, is large:

• Nightly + build-std on every build; per-Rust-version maintenance of a bespoke
  target.
• cfg(unix) flips the entire dep graph, not just perms — signals, process,
  users/groups (getpwuid), termios, mmap, net — much of which calls libc
  symbols WASI libc doesn't implement → link/ENOSYS breakage to chase.
• The fs backend (target_os = "wasi") and MetadataExt (target_family = unix) are gated on different cfgs and don't cleanly compose; effectively a
  std fork.

C. Move the runtime to WASI preview2 / wasi:filesystem (the clean long-term fix)

preview2 carries richer metadata than preview1's filestat. Migrating (or
upstreaming a wasi MetadataExt::mode() backed by an extended filestat) would
let these tools read real bits natively and let us delete all the patches —
without pretending to be unix.

Ask / decision needed

Decide between: (B) invest in a Linux-like custom target + build-std so fs tools
need zero patches, vs. (C) target preview2, vs. (status quo) keep adding small
per-tool host_fs patches. Leaning C long-term; B is the "make it native" ask
but has broad blast radius. Capturing so the tradeoff is explicit before the
patch pile grows.

Related PRs: #268 (filesystem native-parity: wasi-libc + sidecar host_fs),
#269 (coreutils stat/chmod/ls real permission bits).