Re-running sysknife-setup while the daemon runs fails with ETXTBSY and can leave mismatched binaries #563

Re-running sysknife-setup while the user daemon is running fails with ETXTBSY: text file busy and aborts the wizard, possibly after the CLI binary was already replaced.

What happens

packages/setup/install-binary.js:495 writes the binary with fsp.writeFile(destPath, data, { mode: 0o755 }), which opens the existing file with O_TRUNC. Linux refuses that for a file being executed. The catch at 497-500 handles EACCES and EPERM only and rethrows the rest, and the caller at 754-762 prints Install failed: ... and exits 1.

The installed user unit is enabled with Restart=on-failure, so after any successful install the daemon is normally running from ~/.local/bin/sysknife-daemon. Re-running the wizard is the documented upgrade path, so this hits every upgrade.

The primitive reproduces in two lines:

cp /bin/sleep ./fakebin && ./fakebin 5 &
node -e "require('fs').promises.writeFile('./fakebin', Buffer.from('x'), {mode: 0o755}).catch(e => console.log(e.code))"
# ETXTBSY

Fix

Write to destPath + '.new' with mode 0755, then fsp.rename() it over destPath. Rename is atomic and allowed over a running executable: the running process keeps the old inode. Then restart the user unit if it was active, or tell the user to.

Write both binaries to .new first and rename only when both succeeded, so a failure never leaves a CLI and daemon from different releases.

Done when

A test starts a process from a fake binary, re-installs over it through install-binary.js, and asserts the install succeeds and the file on disk is the new content.

Getting started

CONTRIBUTING.md has the build and test commands. No CLA and no copyright waiver. The project is MIT; sign off each commit with git commit -s (the required dco check enforces it since #732). Comment here before you start, so nobody doubles up.