[BUG] 2.1.143 background agents crash exit 1 before init on macOS arm64

Summary

After upgrading from 2.1.142 → 2.1.143 on macOS 26.4.1 (arm64, T6020), every background-spawned agent (both via the native daemon's agent view and via cmux) crashes within 60-200ms with exit 1 before init. Foreground sessions are completely unaffected. Rolling back the symlink to 2.1.142 restores normal behavior immediately without restarting the daemon.

Repro

# Symlink the bad version
ln -sfn ~/.local/share/claude/versions/2.1.143 ~/.local/bin/claude

# Spawn any background agent — instantly crashes
~/.local/bin/claude --print --agent claude "OK"
echo "exit=$?"   # → exit=1, ~60-200ms

The same command works when the symlink points to 2.1.142.

Crash detail (as surfaced by the daemon)

state: crashed
detail: exit 1 before init —
  …at uSK (/\$bunfs/root/src/entrypoints/cli.js:9092:233)
   at async SKO (/\$bunfs/root/src/entrypoints/cli.js:19558:68)
   at async h4O (/\$bunfs/root/src/entrypoints/cli.js:19595:313)
  Bun v1.3.14 (macOS arm64)
backend: daemon
respawnFlags: ["--agent","claude"]

The visible frames are all in the outer wrapper — the real throw is swallowed, so the trace alone can't identify the failing module. The crash happens before any init log line is emitted.

Environment

| Field | Value |
|---|---|
| Claude Code (broken) | 2.1.143 |
| Claude Code (works) | 2.1.142, 2.1.141 |
| Bun bundle | v1.3.14 |
| macOS | 26.4.1 (build 25E253) |
| Kernel | Darwin 25.4.0, xnu-12377.101.15~1/RELEASE_ARM64_T6020 |
| Arch | arm64 |
| Backend exhibiting crash | native daemon agent view, cmux agent view |
| Foreground REPL | works fine on 2.1.143 |

Workaround

rm ~/.local/bin/claude
ln -s ~/.local/share/claude/versions/2.1.142 ~/.local/bin/claude
# next bg-agent spawn picks up the new symlink — no daemon restart needed

Possibly related

• #59561 — same 2.1.143 + Bun 1.3.14 bundle on Windows: Chrome native host crashes with ERR_INVALID_ARG_TYPE binding named pipe. Different platform, but both look like child-process / IPC channel init regressions in the same bundle.
• #50267 — prior bg-subagent regression (2.1.114) shows this surface area has regressed before.

Happy to attach more raw daemon log lines or run \dtruss\ on a re-symlinked 2.1.143 if helpful.