Khorsheed/dsh-ankh-guard

## Core Features
- Gates agent-initiated restarts behind a credential recorded only after build and test checks pass, with the credential bound to the git commit and limited by maxAgeMinutes.
- Runs a preflight dry boot of the target profile composition in a subprocess before stopping the running instance, refusing restart when the composition fails.
- Provides a watchdog through supervise that restarts the instance after exit, runs canary checks, and rolls back to the last known good version after repeated startup failures.
- Uses git-based checkpoints and resets, creating guard-backup-* branch anchors for discarded HEAD and uncommitted changes so recovery does not depend on reflog.
- Restores interrupted root sessions after a planned restart and delivers a restart report to the initiating session through injected followup context.

## Use Cases
- Coding agents that modify a DSH deployment and need to restart the host only after build, test, and composition checks pass.
- Long-running DSH web instances that require supervised restarts with automatic recovery after failed self-modification attempts.
- Development workflows that need git-based rollback points before batches of code changes and verification after restart.

## Technical Details
- The package is a DSH plugin identified as @khorsheed/dsh-ankh-guard and installed through dsh plugin --profile web add from npm or a GitHub mirror.
- CLI commands include verify, record, status, clear, checkpoint, reset, canary, preflight, restart, schedule-exit, and supervise.
- Preflight boots the full plugin tree with the same engine, pins the webserver port to 0, checks client bundle artifacts, then disposes the dry run.
- Restart reporting and interrupted-session resume are injected as plugin-sourced snapshot user messages, with configuration options such as reportRestartContext, resumeInterrupted, and resumeDelayMs.

## Notes
- Git is required because credentials, checkpoints, and rollback are based on commits, and the deployment directory must be a git repository.
- The restart chain requires full-access permissions because it spawns detached processes, kills processes, and binds ports; sandboxed tool runners can block these operations.
- After installation, the first restart must be driven by the CLI or an external supervisor because the running instance has not loaded the plugin and no watchdog exists yet.