The server handled one client at a time, so a client that connected and sent nothing blocked every other client. The request handling is now in handle_client, and the main loop starts one thread for each connection.
The server closes a connection that stays silent for 5 seconds, so its thread ends.
If a client disconnected before its reply, send_response raised OSError and the server ended. The request loop now logs the error and waits for the next client. An error from accept() still ends the server.
An unknown command ended the server loop, so one bad request stopped the service for every client. The server now sends {"status": "ERROR", "message": ...} and continues. A request with no command, or one that is not a JSON object, gets the same reply.
The server tests stopped the server with an unknown command. They now stop it through accept().
bmspy-ups selected its device from the devices that had data at startup, and exited with status 0 if the selection failed. One failed first read ended the monitor or made it select the wrong device.
The server has a new DEVICES command that lists the configured names, and client.list_devices calls it. bmspy-ups selects from that list and then waits for the first data of its device. A wrong or missing --device now exits with status 2. The service file does not restart on that status, because a restart cannot correct it.
An older server ends on the unknown DEVICES command, so update the server first.
A failed JBD read retried forever inside a client request and blocked every other device and client. The server now starts one polling thread for each device, and a request returns the last snapshot. Only the first request after startup waits, until each device completes one read attempt.
One lock makes sure that only one Bluetooth collection runs at a time. A serial BMS retries a failed read after 1 second and keeps its last snapshot for 30 seconds. An error in a collection does not end the polling of that device.
The server tests left main() in a thread after each test. These threads kept their patches active and delayed the suite. Each test now sends an invalid command, which ends the server loop, and then joins the thread. The suite time drops from about 8.5 seconds to about 2 seconds.
Remove one unused line from a test.
tim