Add SplitFlap Gateway as a target for the app

[avandeputte]

SplitFlap Gateway (MQTT) Integration — Change Summary

This adds an option to drive the display through a SplitFlap Gateway
(Waveshare ESP32-S3-RS485-CAN running SplitFlap Gateway firmware at https://github.com/avandeputte/SplitFlapGateway ) over MQTT, as an alternative to a
local serial port. The design goal was to minimize impact on the existing
splitflap-os codebase: the OS interacts with only two topics —
<prefix>/send and <prefix>/rx.

How it works

A new transport class presents the same interface the existing code already
used for the serial port (.write(), .flush(), .reset_input_buffer(),
.in_waiting, .read(), .close()). Because of this, none of the existing
read/write call sites had to change.

• OS → Gateway on <prefix>/send: the exact RS485 frame bytes the old
  code wrote to the serial port (e.g. m05-A\n, m05c\n, m05d\n) are
  published verbatim. The gateway forwards them straight to the bus.
• Gateway → OS on <prefix>/rx: the gateway publishes JSON
  {"ts":...,"wt":"...","command":"<ascii frame>"} for each frame received
  from a module. The transport extracts command, re-appends the newline the
  gateway stripped, and buffers it so the existing calibration / EEPROM-dump
  read loops parse it exactly as they did with serial.

The gateway MQTT connection is independent of the pre-existing
MQTT / Home Assistant integration (which publishes state and accepts control
commands). They use separate clients and settings.

Files changed

New: server/gateway_transport.py

GatewayTransport, a pyserial-compatible facade over MQTT. Touches only
<prefix>/send (publish) and <prefix>/rx (subscribe). Fails fast on an
unreachable broker so the app can fall back to simulation mode, mirroring how
serial failure is handled.

server/app.py

• Import GatewayTransport (graceful fallback if unavailable).
• New settings keys with defaults: connection_type (serial|gateway),
  gateway_broker, gateway_port, gateway_prefix, gateway_user,
  gateway_password.
• _read_config_file(), _get_connection_type(), _get_gateway_config()
  helpers (readable at startup, before load_settings() runs).
• _open_gateway() and _open_connection(); startup now calls
  _open_connection(), which dispatches on connection_type.
• /serial_port now also sets connection_type='serial' when applied.
• New /connection route (GET/POST) to view/switch/configure the connection.
  Password is preserved on blank resubmit and never echoed back in full.
• No existing ser.write() / ser.read() call sites were modified.

server/templates/index.html

• "Serial Port" settings panel replaced by a "Hardware Connection" panel with
  a Connection Type selector toggling between the serial panel and a new
  gateway panel (broker, port, topic prefix, optional username/password).
• Added a clarifying note on the existing "MQTT / Home Assistant" section.

server/static/app.js

• onConnectionTypeChange(), loadConnectionConfig(),
  applyGatewayConnection(); loadSettingsData() now also loads the
  connection config; applying a serial port syncs the type selector.

README.md

• Feature bullet + a "SplitFlap Gateway (MQTT)" section documenting the topic
  contract and environment variables.

Environment variables (optional, override settings.json)

SPLITFLAP_CONNECTION_TYPE=gateway     # 'serial' (default) or 'gateway'
SPLITFLAP_GATEWAY_BROKER=192.168.1.50
SPLITFLAP_GATEWAY_PORT=1883
SPLITFLAP_GATEWAY_PREFIX=splitflap
SPLITFLAP_GATEWAY_USER=               # optional
SPLITFLAP_GATEWAY_PASSWORD=           # optional

Dependencies

paho-mqtt was already present in requirements.txt; no new dependency added.

[csader]

Great work! Duck-typing the serial interface means zero downstream changes, the fail-fast behavior mirrors serial properly, and the password handling is done right.

A few things I'd like addressed before merging:

Silent publish on disconnect (important):
write() doesn't check self._connected before publishing. If the broker disconnects mid-session, paho silently queues messages and the user won't know commands aren't reaching the display. This is especially concerning during calibration/EEPROM writes. At minimum, write() should log a warning or raise when not connected, otherwise command loss is invisible.

Minor items (non-blocking):

• client_id uses second-precision timestamps (int(time.time())). If the app restarts within the same second (crash loop on Pi), you get a duplicate client ID. Consider appending os.getpid() or a short random suffix.

• _read_config_file() is called independently by _get_connection_type(), _get_gateway_config(), and _get_serial_port() at startup — reads settings.json 2-3 times. Not a real problem, just noting it.

• /connection POST doesn't validate the type field. An unknown value silently falls through to the serial branch. A 400 for unrecognized types would be cleaner.

Once the disconnect issue is handled, this is good to go. Since the feature is opt-in (off by default, requires explicit config), I'm comfortable merging to main without a beta release.

[avandeputte]

All 4 items resolved, i think. Tested locally no issues with the changes.

[csader]

All four items addressed cleanly. The warning-not-raise approach in write() is the right call for pyserial compatibility. Good to go.