fix: goose configure exits silently at the model picker

[ahmadelafify]

Fixes #8373.

Summary

On macOS 26.4 (M5 Pro, tested in kitty and Terminal.app), goose configure exits with code 2 right after Model fetch complete prints. The model picker draws one frame and disappears before any keypress is registered.

The cause is a race between cliclack's spinner and the next prompt. spin.stop() flips a flag and returns, but the spinner's render thread keeps writing to the terminal for a few more ms. The next line opens the model picker, which puts stdin in raw mode and reads. Both widgets fight over the terminal and the picker's first read returns Err, which propagates out as exit 2.

Fix: drop the spinner explicitly and yield 50ms so cliclack tears the render thread down before the picker starts reading. The user does not see the delay.

     spin.stop(style("Model fetch complete").green());
+    // cliclack's spinner render thread outlives `stop()` and races the next
+    // prompt's raw-mode stdin read; drop and yield so the picker's
+    // `interact()` doesn't fail before any key is read. See #8373.
+    drop(spin);
+    tokio::time::sleep(std::time::Duration::from_millis(50)).await;

Testing

• cargo build -p goose-cli --bin goose
• goose configure → Claude Code → picker stays open, arrow keys and enter work.

Before:
https://github.com/user-attachments/assets/27d1d5b6-2d94-4f0b-9f0d-33d20fd7a52a

After:
https://github.com/user-attachments/assets/116f6d7c-84e8-4c6e-9845-81049615cc4b

[jh-block]

I don't think the diagnosis of the cause is quite right. In particular, "the spinner's render thread keeps writing to the terminal for a few more ms" cannot be true, because spin.stop() calls finish_with_state(), which calls finish_and_clear(), which joins the render thread. By the time spin.stop() returns, the render thread is no longer running.

The root cause is a bug in the console crate: it calls select() and then treats any interruption as a fatal error and returns Err. We issue SIGKILL to Claude Code and then proceed -- if the Claude Code subprocess exits after select() begins, SIGCHLD interrupts select() and triggers the bug. The sleep masks the problem because SIGCHLD arrives during the sleep, before select() is entered.

The real fix should be in the console crate, so I'll issue a PR there, but a robust workaround here is just to wait for the child process to be reaped before proceeding, instead of issuing the kill signal and immediately continuing without waiting. I'll open a PR here for that.

[jh-block]

Superseded by #9023