fix(cli): macOS force exit cuts off the telemetry flush at exit #1448

Problem

On macOS, the CLI force-exit backstop can stop the telemetry flush that runs at exit. Data that the SDK has not sent yet is lost.

Cause

1. In the finally of packages/cli/src/cli.ts, scheduleForceExit() arms an unref'd 100 ms timer that calls process.exit() (packages/cli/src/lib/force-exit.ts). It runs only on darwin.
2. When the command's work drains, beforeExit fires. createBeforeExitHandler() in packages/cli/src/lib/telemetry.ts ends the session and starts client.flush(3000).
3. The flush opens HTTP requests. These keep the event loop alive, so the unref'd timer can fire.
4. The timer calls process.exit() before the flush completes. The flush is cut off.

The comment at the call site says the timer is a no-op on clean exits. That is true only until beforeExit starts new I/O. After that, the flush itself is the handle that keeps the loop alive.

Impact

Any data still in the SDK buffer when a command ends can be lost on macOS. For example:

• Metrics: the SDK buffers them for up to 5 s, so the exit flush is often the only send.
• Session updates from Sentry.endSession() in the beforeExit handler.
• Client reports and runtime metrics that are sent at shutdown.

Evidence

The Snake leaderboard (#1437) found this problem. A snake.score metric was lost every time the player quit within about 5 s of game over. A trace log showed that the beforeExit flush started and the process exited before the request completed. #1437 works around this only for the score: it calls client.flush() when the game ends (packages/cli/src/lib/games/score.ts).

I saw the loss only for metrics. The impact on sessions and client reports comes from reading the code. I did not reproduce it.

Possible fixes

• Let the telemetry flush delay the force exit. For example, beforeExit could clear or postpone the timer until flush() settles, with a maximum wait.
• Flush telemetry in the finally of cli.ts before scheduleForceExit(), and keep the existing 3 s limit.
• Arm the timer only after the telemetry flush settles.

Any fix must keep the protection from #1237 and #833 against processes that do not exit.