watcher: inotify arm panics on every Linux start (raw syscall checked with the libc errno accessor)

[poelzi]

Problem

FileChangeWatch guards two raw std.os.linux.* syscalls with the wrong errno accessor:

const wd_rc = std.os.linux.inotify_add_watch(...);  // raw syscall: -errno packed into a usize
if (std.posix.errno(wd_rc) != .SUCCESS) return;     // libc accessor: `if (rc == -1) ... else .SUCCESS`
const wd: i32 = @intCast(wd_rc);                    // panics on the negative rc

codedb links libc, so std.posix.errno resolves to std.c.errno, which reports failure only for rc == -1. A usize never equals -1, so the guard always returned .SUCCESS and the negative rc fell straight through to @intCast.

Verified against the pinned toolchain (0.17.0-dev.813+2153f8143):

inotify_add_watch("/tmp/definitely-not-here") ->
  rc = 18446744073709551614 (0xfffffffffffffffe)
  std.posix.errno(rc)     = SUCCESS   <-- wrong
  std.os.linux.errno(rc)  = NOENT     <-- correct

Why it fires for everyone on Linux

armInotify watches /tmp/codedb-notify — the optional muonry interop file, which codedb itself never creates. On any machine without muonry that watch returns -ENOENT on every arm, so:

• Debug / ReleaseSafe — hard panic (integer does not fit in destination type) in the watcher thread.
• ReleaseFast — @intCast is unchecked, so the value truncates to a garbage watch descriptor instead. That is why shipped releases never surfaced it: /tmp/codedb-notify also fails silently, and a failing directory watch (EACCES, watch-limit ENOSPC) inserts a bogus wd -> rel mapping that can misattribute later events.

The kqueue path already tolerated the file's absence (else |_| {}); only inotify did not.

Failing test

Reproduces today on release/0.2.5842 as a crash in the existing test:

$ zig build test-index -Dtest-filter="issue-690"
thread panic: integer does not fit in destination type
src/watcher.zig:1494:25: in addInotifyWatch
src/watcher.zig:1482:29: in armInotify
src/watcher.zig:1336:28: in arm
src/watcher.zig:1621:14: in incrementalLoop
Build Summary: 2/3 tests passed (1 crashed)

Fix

Classify raw returns by the -errno range instead. rawLinuxSyscallFailed mirrors std.os.linux.errno without referencing the linux-only namespace, so the boundary behaviour stays unit-testable on every target. Applied at both call sites — inotify_init1 had the identical bug.

I scanned src/ for other raw std.os.linux.* call sites; these two are the only ones.

Verification

• New test watcher: raw linux syscall failures are detected by the -errno range pins success values, -ENOENT, and the -1 / -4095 / -4096 boundaries.
• Both tests confirmed to actually catch the bug: with the old classification restored, issue-690 crashes and the new test fails; with the fix, both pass.
• Full suite green: zig build test 25/25 steps, 807/811 passed, 4 skipped, 0 failed, 0 crashed (previously 1 crash).
• Daemon smoke-tested with /tmp/codedb-notify absent — arms and runs clean.