fix: Don't use eprintln! in the panic hook

[molofsky]

Problem

panic_handler ends with eprintln!, which panics if the write to stderr fails. A panic raised inside a panic hook aborts the process immediately, so the hook never completes and the crash report it exists to print is lost. The user sees a bare Aborted (core dumped) with no diagnostics at all.

This is reachable in practice during TUI teardown, where the terminal fd can already be gone. Observed on 2.10.1 with a locally symbolized build:

turborepo_ui::tui::app::cleanup              crates/turborepo-ui/src/tui/app.rs:1111
  -> ratatui Terminal::drop                  ratatui-core/src/terminal/terminal.rs:119
       eprintln!("Failed to show the cursor") -> write fails -> panic
  -> turborepo_lib::panic_handler            crates/turborepo-lib/src/panic_handler.rs:66
       eprintln!  -> write fails -> panic
  -> double panic -> abort()  (SIGABRT)

ratatui's Drop reports a failed show_cursor via eprintln!; that write fails too and panics; the hook then panics the same way, and the second panic aborts. The first panic is ratatui's to own — but the hook shouldn't be capable of turning any panic into an abort.

Fix

Write to stderr directly and discard the error:

let _ = std::io::stderr().write_all(
    format!("Oops! Turbo has crashed.\n\n{report_message}\n").as_bytes(),
);

A panic hook runs precisely when things are already broken, so it shouldn't assume its own output path works. If stderr is unusable we now print nothing, instead of escalating to an abort. When stderr is usable — the overwhelmingly common case — behaviour is unchanged.

Notes / testing

• Applies to main; eprintln! is the only panicking macro in the hook.
• The equivalent change was built and exercised on 2.10.1 (Linux, musl target): the double-panic abort stopped reproducing.
• The changed code compiles standalone under -D warnings (no unused import, no type error).
• I have not built the full workspace against main — I don't have capnp/zig and the pinned nightly on this machine. Happy for CI to be the judge, which is why this is a draft.

Filed as a draft for maintainer review; glad to adjust the wording or add a regression test if useful.

[vercel]

Someone is attempting to deploy a commit to the Internal Apps Team on Vercel.

A member of the Team first needs to authorize it.

[anthonyshew]

Thank you!