[MattieGit]

Summary

Connection handling, write safety and request batching (v1.15.0).

• Lost link detected. The connected flag was only cleared by close(), so a dropped link was never noticed and the reconnect backoff never ran. Link failures (ConnectionException, ModbusIOException, OSError) now mark the client disconnected and close the pymodbus transport; the next poll reconnects.
• Dead device stops the poll early. After a failed block, the first single read acts as a probe; if it also hits a link failure the poll stops. get_all_data() returns None when the link is lost mid-poll. Modbus error replies keep the per-entity fallback.
• Warn-once re-arms when a target reads fine again, so a later outage is logged at WARNING again.
• Setpoint validation. EntityDef gains optional min_value/max_value; the four writable setpoints use the ranges both HA integrations already expose (DHW 40–65, heating 20–65, cooling 7–25 °C). NaN, inf and out-of-range values are refused before any Modbus traffic. 16-bit writes round instead of truncating and refuse values that do not fit.
• SG Ready set atomically with one write_coils request (falls back to two single writes if the device rejects it); get_sg_ready_mode() reads both coils in one request.
• read_all_* batched: 94 → 9 requests, same keys, key order and values.
• async_get_mac_address() no longer blocks the event loop (DNS lookup and /proc/net/arp read run in a thread).

Measurement	Before	After
Requests per get_all_data() poll, device unreachable	68	2
Poll on a device that accepts TCP but never answers (real pymodbus)	96 s, reports connected	24 s, returns None, reconnects next poll
read_all_sensors + read_all_binary_sensors + read_all_switches	94	9

Compatibility

• Public API unchanged; new EntityDef fields are optional and appended at the end.
• get_all_data() returning None on a lost link is within its documented contract; the HA core coordinator already raises UpdateFailed on None.
• Values the integrations' UIs allow are all accepted; out-of-range values sent through the HACS write_register service are now refused.
• Multi-coil write (FC15) is untested on a real Qube; the single-write fallback covers a device that rejects it.

Test plan

• Library: 130 passed (47 new), ruff check and ruff format --check clean
• HA core tests/components/hr_energy_qube against this branch: 30 passed (same as 1.12.0 and main)
• HACS integration tests against this branch: 332 passed (same as main)
• End-to-end against a real pymodbus TCP server: decoding, SG Ready write/read, setpoint write/read-back and range refusal, server stop → None without delay, restart → reconnect, silent device → 24 s
• SG Ready mode change on real Qube hardware