The fleet picks each member's port at start, but llama-swap only spawns
the member on its first request, so the port sits unclaimed for the
whole lazy wait and a passing ephemeral connection can take it. The
server then exits "couldn't bind HTTP server socket", llama-swap masks
it as "exited prematurely", and the client classified every such death
as CONNECTION: the sole replica got written off as unhealthy and the
request failed with "no healthy replica", even though llama-swap
re-spawns on the next request and normally binds fine (the occupier is
gone within seconds).

When the upstream's captured tail shows the bind failure, classify the
death as SERVER instead, so the existing retry_on_busy backoff re-drives
the spawn. Any other exit reason keeps the CONNECTION failover path.