Let the standalone installers start stacktape init [skip ci] - #318

The standalone installers can now start stacktape init right after installing, so a machine without Node.js needs one command:

curl -fsSL https://installs.stacktape.com/macos-arm.sh | sh -s -- init

iex "& { $(irm https://installs.stacktape.com/windows.ps1) } init"

Anything after the installer's own arguments is passed to the installed CLI, run by its full path because the new PATH entry only applies to new shells. With curl | sh, the CLI's input comes from the terminal (/dev/tty), because stdin is the rest of the piped script. Without a terminal (CI, Docker builds, a coding agent's shell), the CLI gets empty input instead of failing, so sh -s -- init --headless works there too. The shell installers exec the CLI, so its exit code is the installer's. Run as a file (powershell -File windows.ps1 init), the Windows installer exits with the CLI's code; under iex it never calls exit, which would close the user's PowerShell window. Without extra arguments nothing changes. The docs homepage's standalone-install tabs now use these commands.

Testing the real flow found a separate bug: released binaries never found the init wizard beside them, so stacktape init said "The wizard interface has not been built" and served a blank page. This affects 4.0.0-preview.10. findWizardBundle looked next to process.argv[1], which is a virtual /$bunfs/... path in a Bun-compiled binary. It now uses process.execPath.

How it was checked:

• installers.spec.ts pipes each shell installer into sh -s -- init --projectDirectory "my app" with no terminal (setsid) and inside a pseudo-terminal (script). It checks the exact arguments, the CLI's stdin and the exit code. windows.spec.ts runs both Windows installers under pwsh as a file and through the iex form. The new cases fail against main's installers.
• End to end on Linux: I built a release archive from this branch and installed it through the piped linux.sh in a pseudo-terminal. init started; headless Chromium rendered the wizard; Ctrl+C ended it. init --headless wrote stacktape.yml.
• End to end on Windows 11: I installed the branch's windows.zip with windows.ps1 run as a file in Windows PowerShell 5.1 and pwsh 7, using temporary home folders. init --headless --projectDirectory "C:\…\my app" wrote stacktape.yml, and the installed stacktape.exe served the wizard page. I did not run the irm | iex download form on Windows: Windows Defender flags a powershell.exe command line that downloads and runs a script as Trojan:Win32/Commando.A!ml. That form was checked with pwsh on Linux instead.
• pnpm check:public passes except two failures that origin/main has too: scripts/perf/packaging-harness.spec.ts ("measures a shape and reports which artifacts each edit changes") and the test registry, which finds no owner for apps/cli/src/domain/packaging-manager/dependency-installer.spec.ts from Reuse the runner's dependency install on Git push deploys [skip ci] #316. Also shellcheck (no new findings), and independent reviews by Fable 5.1 and GPT-6.1 Sol. Sol found that the file-run Windows installer hid the CLI's exit code, which is now fixed.

Nothing is published. The release workflow's publish-installers job uploads the installers to installs.stacktape.com on the next stable release. Until then the live 3.7.1 installers ignore init and only install. Console PR stacktape/console-app#142 records that the docs should be published after that release.

Two unverified notes from review. In Windows PowerShell 5.1 with $ErrorActionPreference = 'Stop', a CI step that redirects stderr could show an error after the CLI has finished. A coding-agent shell on macOS or Windows that runs init without --headless waits for the browser wizard, as plain stacktape init does there.