fix(esp32): reboot out of a wedged network stack instead of needing a power cycle (#106)

Node 35 dropped off the fabric and stayed off for 27 min until it was power-cycled
by hand. The chip was fine (bootreason power-on, RS-485 polling clean throughout) --
nothing in the firmware watches the network, since the only WDT-subscribed task talks
to the A/C bus, not Wi-Fi.

- net-loss reboot watchdog (15 min, ESPHome's reboot_timeout default). Liveness =
  associated AND link-local IPv6 valid; Matter is IPv6-only and esp-idf#5796 can kill
  the LL address while Wi-Fi/IPv4 stay healthy. Gated against firing mid-OTA.
- force WIFI_PS_NONE (IDF defaults to modem sleep), falling back to MIN_MODEM while a
  BLE window is open for single-radio coex.
- feed the task WDT per drained TX item: a full queue drain on a silent bus exceeded
  the 5 s timeout and could panic-reboot the node on its own.
- CONFIG_HISENSE_RS485_HW_MODE (default off): opt-in native UART_MODE_RS485_HALF_DUPLEX
  DE handling, matching ESPHome's uart component, for bench comparison.

Deployed and verified on node 35 as esp32 v1.1.12 (int 10112). AmebaZ2 version bumped
to 1.3.28 because the shared driver changed (the new feed is #ifdef ESP_PLATFORM, so
the AmebaZ2 binary is functionally unchanged).

Assisted-by: AI

Author: AndrewDemsDS

firmware/esp32-matter/CMakeLists.txt

@@ -2,7 +2,7 @@
 #   . ~/esp/esp-idf-v5.5.4/export.sh && . $ESP_MATTER_PATH/export.sh
 #   idf.py set-target esp32 && idf.py build flash monitor
 cmake_minimum_required(VERSION 3.16)
-set(PROJECT_VER "1.1.11")
+set(PROJECT_VER "1.1.12")
 
 # Unified versioning (issue #77): the Matter softwareVersion INT is DERIVED from PROJECT_VER --
 # MAJOR*10000+MINOR*100+PATCH -> a readable, strictly-monotonic uint32. This keeps the human

firmware/esp32-matter/components/hisense_hal/hisense_hal.c

@@ -81,12 +81,27 @@ void serial_baud(serial_t *obj, int baud)
         .source_clk = UART_SCLK_DEFAULT,
     };
     uart_param_config(HAL_UART_NUM, &cfg);
+#ifdef CONFIG_HISENSE_RS485_HW_MODE
+    /* Opt-in hardware half-duplex (see CONFIG_HISENSE_RS485_HW_MODE's help text). The DE line
+     * becomes this UART's RTS, and the peripheral asserts it exactly while the TX FIFO drains --
+     * the same sequence ESPHome's uart component uses for every RS-485 device it supports:
+     *   uart_set_pin(num, tx, rx, flow_control_pin as RTS, UART_PIN_NO_CHANGE)
+     *   uart_set_mode(num, UART_MODE_RS485_HALF_DUPLEX)
+     * PA_17 is the DE GPIO (PinNames.h remaps it to the ESP32 pin actually wired). */
+    uart_set_pin(HAL_UART_NUM, obj->tx, obj->rx, PA_17, UART_PIN_NO_CHANGE);
+#else
     uart_set_pin(HAL_UART_NUM, obj->tx, obj->rx,
                  UART_PIN_NO_CHANGE, UART_PIN_NO_CHANGE);
+#endif
     if (!s_evt_q) {
         uart_driver_install(HAL_UART_NUM, HAL_UART_RX_BUF, 0,
                             HAL_UART_EVT_QLEN, &s_evt_q, 0);
     }
+#ifdef CONFIG_HISENSE_RS485_HW_MODE
+    /* Order matters: ESP-IDF requires uart_set_mode() AFTER uart_driver_install(), because the
+     * install resets the RS485 config registers. ESPHome sequences it the same way and says so. */
+    uart_set_mode(HAL_UART_NUM, UART_MODE_RS485_HALF_DUPLEX);
+#endif
 #ifdef HISENSE_HAL_UART_INTERNAL_LOOPBACK
     // DIAGNOSTIC ONLY: internally wire this UART's TX to its own RX inside the
     // chip, so a transmit must show up as a receive if the RX path is healthy.
@@ -145,11 +160,19 @@ void serial_putc(serial_t *obj, int c)
     uint8_t b = (uint8_t)c;
     uart_write_bytes(HAL_UART_NUM, &b, 1);
     g_hal_tx_bytes++;
+#ifndef CONFIG_HISENSE_RS485_HW_MODE
     // Block until the byte has actually left the wire. uart_write_bytes() only
     // queues to the TX FIFO; hisense_tx_raw() drops the DE line right after its
     // putc loop, so without this the tail of the frame would be cut off while DE
     // is already low (RS-485 half-duplex). At 9600 baud this is ~1ms/byte.
     uart_wait_tx_done(HAL_UART_NUM, pdMS_TO_TICKS(20));
+#else
+    // Hardware RS-485 mode: the peripheral holds DE asserted until the FIFO and
+    // shift register are empty, so there is nothing to race and no reason to
+    // block here. Letting the bytes accumulate in the FIFO is also what makes
+    // this a one-shot frame write (ESPHome's write_array) rather than 16
+    // separate stop-and-wait transmissions.
+#endif
 }
 
 void serial_free(serial_t *obj)

firmware/esp32-matter/main/Kconfig.projbuild

@@ -20,4 +20,33 @@ config HISENSE_DEBUG_BUILD
         flavours. Anything that changes bus or Matter behaviour would make bench
         results meaningless for the shipped image.
 
+config HISENSE_RS485_HW_MODE
+    bool "RS-485: let the UART peripheral drive DE (hardware half-duplex)"
+    default n
+    help
+        UNVALIDATED, BENCH-ONLY. Leave this OFF unless you are actively comparing
+        the two paths on a debug node with a scope or a frame-error count.
+
+        Default (OFF) is the path this project validated on real hardware: the
+        shared driver toggles the DE line itself as a plain GPIO, waits
+        HISENSE_DE_SETTLE_MS before the first byte, and holds DE high for
+        HISENSE_DE_DRAIN_MS after the last so the FIFO and shift register empty
+        before the transceiver stops driving. Finding that DE existed at all cost
+        a multi-day debug, so the default does not change.
+
+        When ON, DE is handed to the UART peripheral instead: the DE GPIO becomes
+        UART RTS and the port is put into UART_MODE_RS485_HALF_DUPLEX, so the
+        hardware asserts DE exactly while the TX FIFO drains and releases it on
+        the same clock as the bit stream. The software toggle and both delays are
+        compiled out, and serial_putc() stops waiting per byte.
+
+        This is what ESPHome's uart component does for every RS-485 device it
+        supports (flow_control_pin -> uart_set_pin RTS -> uart_set_mode), so it is
+        well proven in general -- just not yet against THIS A/C. It cannot desync
+        the way a software toggle can, and it removes ~30 ms of dead time per
+        frame.
+
+        AmebaZ2 is unaffected either way: it has no such peripheral mode and keeps
+        the software path unconditionally.
+
 endmenu