fix: wipe migrated config files too when clearing dead registrations

[compscidr]

Problem

bump-android-nas.local got stuck in an unrecoverable crash loop after the runner registration it had persisted (agentId=1891) was pruned server-side by GitHub while the host was offline.

The preflight from #77 correctly detected the 404 and wiped .runner, .credentials, and .credentials_rsaparams — but the persist dir also held a .runner_migrated (written by the runner's broker migration), which the wipe doesn't know about. The runner's IsConfigured() treats the migrated file as a valid config (ConfigurationStore.cs):

bool configured = new FileInfo(_configFilePath).Exists || new FileInfo(_migratedConfigFilePath).Exists;

So config.sh refused to register ("already configured") while run.sh crashed loading the missing .runner (Value cannot be null. (Parameter 'configuredSettings')) — looping forever. The preflight can't recover either, since it gates on .runner existing.

Fix

• The wipe now also removes .runner_migrated and .credentials_migrated (IsConfigured()/HasCredentials() treat them as equivalent to the primary files).
• The wipe also cleans the same files from /actions-runner: the container's writable layer survives restarts, and the upstream entrypoint's persist-back step re-seeds the persist dir from it (observed live — deleting the file from the volume alone got it immediately re-created on the next loop iteration).
• A preflight pass heals the already-wedged state (.runner_migrated present without .runner) left behind by images with the old wipe, so affected hosts recover on their own once they pull this image.

Testing

• bash -n + shellcheck clean.
• Script-level test of the three paths: wedged state gets healed, empty persist dir is a no-op, healthy .runner without ACCESS_TOKEN is left untouched.
• The equivalent manual cleanup (removing .runner_migrated from both the volume and /actions-runner) unstuck the real crash-looping runner on nas.local, which registered fresh and immediately picked up a job.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR hardens the runner-registration preflight so it can recover from a “wedged” state caused by GitHub Actions runner broker migration files (.runner_migrated / .credentials_migrated) surviving a registration wipe, which can otherwise block re-registration while still causing run.sh to crash-loop.

Changes:

• Add a centralized wipe_registration() helper that removes both primary and migrated config/credential files.
• Extend wipe behavior to also remove the same files from /actions-runner to prevent the upstream persist-back step from re-seeding the volume.
• Add a preflight healing step that detects and fixes the specific orphaned migrated-config state (.runner_migrated present without .runner).

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.