Register Falcon cluster endpoints with Envoy

[samuel-williams-shopify]

Summary

• Add a supervised Falcon cluster adapter using prepare_worker!(instance, listener:).
• Convert Falcon listeners into Envoy-specific endpoint / endpoints supervisor state.
• Preserve the endpoint scheme and supported wire-protocol names in supervisor state.
• Intersect worker protocol capabilities per cluster and select the preferred common Envoy upstream mode.
• Publish grouped IP and Unix-domain-socket addresses through EDS.
• Add a two-worker Falcon Unix-socket example.

Falcon retains its server-side Listener terminology. At the integration boundary, supervisor-envoy converts that value into an upstream Envoy Endpoint, avoiding confusion with Envoy LDS listeners. Protocol implementation objects do not cross the supervisor-state boundary.

Released dependencies

• async-http v0.97.0
• async-service-supervisor v0.18.0

Pending dependencies

• Add Falcon cluster service environment falcon#356
• Support grouped Unix endpoint addresses async-grpc-xds#1

Testing

• RUBYLIB=../async-grpc-xds/lib bundle exec bake test (21 tests, 35 assertions)
• bundle exec rubocop
• bundle exec bake decode:index:coverage lib (26/26 definitions)

[samuel-williams-shopify]

Local cross-repository validation completed with the Falcon, supervisor, xDS, and supervisor-envoy branches wired through RUBYLIB:

• Started examples/falcon/falcon.rb with two real cluster workers.
• Confirmed both generated Unix sockets accepted HTTP requests.
• Connected an ADS client to the supervisor xDS server.
• Confirmed EDS contained two LbEndpoint resources whose primary addresses were the exact worker pipe.path values.

I also attempted the final request through a real Envoy container. The local Colima/containerd runtime is currently unhealthy: multiple pre-existing and new container tasks are stuck in UNKNOWN, and Docker cannot start or remove containers without hanging. I did not restart the user-wide VM. The Ruby-side listener -> endpoint state -> EDS pipe path chain is confirmed; the final Envoy proxy hop should be rerun after Colima is repaired.

Merge/release sequencing matters for standalone CI: this PR needs socketry/async-service-supervisor#7 and socketry/async-grpc-xds#1 (and the current xDS control-plane API) available before its normal bundle can load the new APIs.