Per-connection supervisor using :rest_for_one strategy.

Manages a Finch pool (unless the config names an existing one with :finch_name) and a BatchWriter GenServer for a single named InfluxDB connection with crash isolation.

On start, registers the connection the client initialised in InfluxElixir.Connection so it can be resolved by name from any process via Connection.fetch!/1 or the facade. The registration is taken down, and the client's state for the connection released, whenever the supervisor stops, including when it gives up after too many restarts or a child fails to start: a name never resolves to a connection that is not running. A call made in the moment between the registration and the pool starting is {:error, {:connection_error, :no_pool}}, never a raise.

If the Finch pool crashes, the connection's BatchWriter (at most one) restarts with it, since it depends on the pool. A BatchWriter crash does NOT take down the pool. The writer is started with the connection's name (connection: name), so it writes through whatever is registered and its child spec, which supervisors print when a start fails, holds no token.

Children stop in reverse start order, so on shutdown the BatchWriter writes its buffer while the Finch pool is still up.

If this supervisor is killed outright, its children outlive it for a moment (they stop when they see it exit, the writer after writing its buffer) and still hold their registered names. A restart waits for those names to be released, up to 15000 ms each, before starting new children: starting at once failed with :already_started, and the repeated failures took the top-level supervisor, and every other connection, down with it.

Failures to start

A client whose init_connection/1 returns {:error, reason} makes start_link/1 return {:error, reason}. The :token never appears in that error or in the logs of a failed start: the child spec and the supervisor's argument carry it opaque (inspect/1 prints #InfluxElixir.ConnectionSupervisor.Secret<...>).