fix: keep the Homebrew symlink in the launchd plist

[vyncint]

Found while verifying the v0.2.8 cask install end to end on the reference machine.

The plist pinned a version directory

setup resolves symlinks before recording the driver's path, which on a Homebrew install produced:

setup: driver installed as a launchd service … binary=/opt/homebrew/Caskroom/openshell-driver-applecontainer/0.2.8/openshell-driver-applecontainer

That path belongs to one version. The plist now keeps Homebrew's <prefix>/bin symlink, which is stable across versions; every other symlink is still resolved, so the plist does not depend on one staying put. Verified live — setup on a cask install now records /opt/homebrew/bin/openshell-driver-applecontainer.

The Caskroom detection moved to internal/hostsetup and is shared with update, which needs the same answer for the opposite reason (it must not write into the Caskroom).

What testing the upgrade path actually revealed

I first added a postflight launchctl kickstart so a bare brew upgrade would restart the driver on the new binary. Testing it showed the hook can never fire, and why:

brew upgrade removes the driver's launchd service. Homebrew replaces a cask by uninstalling the old version first, and that runs the cask's uninstall launchctl: directive. From cask/artifact/uninstall.rb:

UPGRADE_REINSTALL_SKIP_DIRECTIVES = [:signal].freeze

Only signal is skipped on upgrade/reinstall — launchctl always runs. Confirmed on the machine: after brew reinstall --cask, the agent was gone and ~/Library/LaunchAgents/local.openshell-driver-applecontainer.plist was deleted. The postflight kickstart then had nothing to restart, so it is dropped rather than shipped as dead code.

Keeping the uninstall launchctl: directive is deliberate: dropping it would preserve the service across upgrades, but leave a real brew uninstall with a loaded agent respawning a binary that no longer exists.

So setup after an upgrade is mandatory, not advisory. The README and the cask caveats now say that plainly instead of the weaker "re-run setup so the service restarts on the new binary".

Verification

On the reference machine, through a local tap serving a snapshot build:

• setup on a cask install records the <prefix>/bin symlink (was: the Caskroom path)
• cask install with no agent present → postflight succeeds, install completes
• brew reinstall --cask → agent unloaded and plist deleted, confirming the upgrade behaviour above
• machine restored to the real v0.2.8 cask and re-setup afterwards

go test ./..., make lint, goreleaser check, and ruby -c on the generated cask all clean. Unit tests cover the cask symlink, an ordinary symlink, and an unresolvable path.

Note: the earlier test (ubuntu-latest) failure on this branch was TestPollerTracksExit in internal/grpcsvc — a timing-sensitive poller test untouched by this change. It passed on rerun and 8 local -race runs; tracked separately.