Starts Elixir's Registry for the :local backend, tolerating a restart that races the previous registry's own shutdown.

Why this exists

Registry.start_link/1 starts a Registry.Supervisor registered under the given name, and that supervisor's children are Registry.Partition processes registered under derived names such as Sagents.Registry.PIDPartition0.

Registry.Partition traps exits, because it links to every registered process. So when Sagents.Registry dies abnormally, the partition does not die with it: it receives the exit signal as a message and terminates asynchronously, holding its registered name for a few milliseconds longer.

Sagents.Supervisor restarts its failed child immediately, well inside that window. Registry.Supervisor.init/1 then tries to start a partition under a name the outgoing one still holds and gets {:already_started, pid}. The restart fails, and because a failed restart is not something a supervisor retries its way out of, Sagents.Supervisor gives up and exits :shutdown — taking the registry, both dynamic supervisors and every running agent with it, permanently. Nothing below it comes back until the host application's own supervisor restarts the whole subtree.

This module closes that window: on an :already_started collision it waits for the straggler to actually exit, then retries. The wait is a monitor, not a sleep, so it costs exactly as long as the collision lasts.

Only the :local backend needs this. Under :horde the name Sagents.Registry belongs to Horde.RegistryImpl, which Horde's own supervisor restarts internally, so there is no name for a restart to collide with.

This covers the case where Registry.Supervisor itself dies. The other :local failure mode, where the supervisor survives and its partition is replaced with empty tables, produces no failed child at all and is handled by Sagents.RegistryWatcher.