Sync pipeline state branch to remote for durability

[james-in-a-box]

The orchestrator persists pipeline state on a local-only git branch (egg/pipeline-state) backed by a Docker named volume. This state is never synced to remote, meaning:

• Volume loss = total state loss — if the Docker volume is deleted or corrupted, all pipeline history is gone
• No cross-host recovery — spinning up the orchestrator on a different host starts with a blank slate
• Old pipelines can't be inspected externally — unlike checkpoints (which are pushed to egg/checkpoints/v2), pipeline state is invisible outside the container

Additionally, on startup reconciliation, old pipeline state files with outdated schemas (e.g., removed agent roles like reviewer_unified) cause validation errors that log warnings on every restart indefinitely since there's no migration path or cleanup mechanism.

Proposal:

1. Push the egg/pipeline-state branch to remote (similar to how egg/checkpoints/v2 works)
2. Pull on startup to restore state on fresh volumes/hosts
3. Consider a migration or schema-compat layer for old pipeline state files that don't validate against the current Pydantic model (e.g., skip or auto-migrate unknown enum values)

— Authored by egg