touch: beta_32 — non-blocking TCP/WS companion writes (Wi-Fi freeze fix) + chunked history flush + stall diagnostics

Root cause of the beta_31 periodic 1-2s freezes with Wi-Fi on: every
received mesh packet pushes an RX-log frame to all connected TCP/WS
companion clients from the loop thread, and writeAllBytes called
WiFiClient::write() unguarded — against a stale peer (phone asleep,
half-open socket) that blocks in the framework's 1s select() retries.
The 120ms app timeout only bounded attempts, not the first call.

- TCP+WS servers: probe writability (zero-timeout select on the client
  fd) before every write; lwIP reports writable only with ~2.8KB free,
  a whole frame is ~180B, so probed writes complete instantly.
- Reap clients whose writes fail continuously for 10s (half-open
  sockets previously stayed "connected" forever, eating the write
  budget on every pushed frame).
- Chat-history ring: chunked writer + wider flush spacing (same
  treatment saveContacts got in beta_30).
- Stall tracer: 16-entry ring of >=0.2s loop stalls, shown in About.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

deploy/meshamerica-catalog.json

@@ -17,19 +17,19 @@
           "role": "gui",
           "title": "Touch UI",
           "version": {
-            "beta_27": {
-              "notes": "Fixed a reboot loop that could hit while the device saved its contact list. Internal-flash housekeeping during a save could freeze the device long enough to trip the watchdog and restart it; that save no longer trips it. Thanks to the crash report that pinned it down.\nDevices installed under the Launcher no longer show the on-device \"Install update\" button. That update path only works on a standalone flash; under the Launcher it failed with \"won't fit: Bad Size Given\". The \"update available\" note stays, but you update through the Launcher instead. Thanks to Keith (KM1USR) for the report.\nKeyboard backlight brightness control on the T-Deck: a live slider plus Off / 25 / 50 / 75 / Max presets in Settings > Keyboard. It dims smoothly (fine-grained at the low end), applies to the on and auto modes and the new-message flash, and survives a reboot. Thanks to Yazutsu (PR #85).\nOn devices with a microSD card, contacts and channels are now kept on the card (like chat history and map tiles already are). Your existing contacts move over automatically on first boot; your identity and settings stay in internal flash, so the node is unchanged. Keep the card inserted afterwards - the contact list lives on it, and an older firmware won't look there if you ever downgrade.\nThis is the new stable build, promoted from the beta_27 test channel. It includes everything from the beta_26 and beta_27 test builds.",
+            "beta_31": {
+              "notes": "The chats tab has a fresh look: each conversation now shows a round avatar in that chat's signature colour (the same colour its bubbles use) with the first two letters of the name, the last message previewed under the name, the time top right and the unread count below it. Compact chat mode keeps the dense single-line rows instead, so both styles are available.\nChannels can have a custom icon: open the channel's settings (long-press or the gear) and tap \"Chat icon\" to pick any emoji as its avatar. Long-press the same entry to go back to the automatic letters. The choice is remembered per channel.\nLong chat names no longer wrap onto the message preview underneath.",
               "files": [
                 {
                   "type": "flash-wipe",
                   "name": "wadamesh-heltec-v4-tft-merged.bin",
-                  "url": "https://firmware.wadamesh.com/releases/TOUCH/beta_27/wadamesh-heltec-v4-tft-merged.bin",
+                  "url": "https://firmware.wadamesh.com/releases/TOUCH/beta_31/wadamesh-heltec-v4-tft-merged.bin",
                   "title": "Full install (bootloader + firmware)"
                 },
                 {
                   "type": "flash-update",
                   "name": "wadamesh-heltec-v4-tft.bin",
-                  "url": "https://firmware.wadamesh.com/releases/TOUCH/beta_27/wadamesh-heltec-v4-tft.bin",
+                  "url": "https://firmware.wadamesh.com/releases/TOUCH/beta_31/wadamesh-heltec-v4-tft.bin",
                   "title": "Update (app only)"
                 }
               ]
@@ -49,19 +49,19 @@
           "role": "gui",
           "title": "Touch UI",
           "version": {
-            "beta_27": {
-              "notes": "Fixed a reboot loop that could hit while the device saved its contact list. Internal-flash housekeeping during a save could freeze the device long enough to trip the watchdog and restart it; that save no longer trips it. Thanks to the crash report that pinned it down.\nDevices installed under the Launcher no longer show the on-device \"Install update\" button. That update path only works on a standalone flash; under the Launcher it failed with \"won't fit: Bad Size Given\". The \"update available\" note stays, but you update through the Launcher instead. Thanks to Keith (KM1USR) for the report.\nKeyboard backlight brightness control on the T-Deck: a live slider plus Off / 25 / 50 / 75 / Max presets in Settings > Keyboard. It dims smoothly (fine-grained at the low end), applies to the on and auto modes and the new-message flash, and survives a reboot. Thanks to Yazutsu (PR #85).\nOn devices with a microSD card, contacts and channels are now kept on the card (like chat history and map tiles already are). Your existing contacts move over automatically on first boot; your identity and settings stay in internal flash, so the node is unchanged. Keep the card inserted afterwards - the contact list lives on it, and an older firmware won't look there if you ever downgrade.\nThis is the new stable build, promoted from the beta_27 test channel. It includes everything from the beta_26 and beta_27 test builds.",
+            "beta_31": {
+              "notes": "The chats tab has a fresh look: each conversation now shows a round avatar in that chat's signature colour (the same colour its bubbles use) with the first two letters of the name, the last message previewed under the name, the time top right and the unread count below it. Compact chat mode keeps the dense single-line rows instead, so both styles are available.\nChannels can have a custom icon: open the channel's settings (long-press or the gear) and tap \"Chat icon\" to pick any emoji as its avatar. Long-press the same entry to go back to the automatic letters. The choice is remembered per channel.\nLong chat names no longer wrap onto the message preview underneath.",
               "files": [
                 {
                   "type": "flash-wipe",
                   "name": "wadamesh-tdeck-merged.bin",
-                  "url": "https://firmware.wadamesh.com/releases/TOUCH/beta_27/wadamesh-tdeck-merged.bin",
+                  "url": "https://firmware.wadamesh.com/releases/TOUCH/beta_31/wadamesh-tdeck-merged.bin",
                   "title": "Full install (bootloader + firmware)"
                 },
                 {
                   "type": "flash-update",
                   "name": "wadamesh-tdeck.bin",
-                  "url": "https://firmware.wadamesh.com/releases/TOUCH/beta_27/wadamesh-tdeck.bin",
+                  "url": "https://firmware.wadamesh.com/releases/TOUCH/beta_31/wadamesh-tdeck.bin",
                   "title": "Update (app only)"
                 }
               ]

release-notes/beta_32.txt

@@ -0,0 +1,9 @@
+# beta_32 (TEST) — the Wi-Fi freeze fix: a stale phone/web connection can no longer stall the device. On top of beta_31.
+# One user-facing note per non-blank, non-# line; # lines are section comments.
+
+# --- Fixed ---
+The 1-2 second freezes every ~10 seconds some users saw with Wi-Fi on are gone. When a phone app or web client was connected over Wi-Fi and that connection went stale (phone asleep, walked out of range), every received mesh packet made the device wait on the dead connection before the UI could continue. The device now only writes to a connection when it can accept data immediately, and a connection that stays unresponsive for 10 seconds straight is dropped so the client can reconnect cleanly.
+Saving chat history could briefly hitch the UI on busy meshes: the history file now writes in a few large chunks instead of thousands of tiny ones (the same treatment the contact list got in beta_30), and background saves are spaced further apart while messages stream in.
+
+# --- New ---
+The About page now shows recent loop stalls (anything over 0.2 seconds, newest first), so freeze reports can be confirmed or ruled out at a glance.

src/helpers/esp32/TCPCompanionServer.cpp

@@ -1,20 +1,44 @@
 #include "TCPCompanionServer.h"
 
+#include <lwip/sockets.h>
+
 #define RECV_STATE_IDLE        0
 #define RECV_STATE_HDR_FOUND   1
 #define RECV_STATE_LEN1_FOUND  2
 #define RECV_STATE_LEN2_FOUND  3
 #define TCP_WRITE_TIMEOUT_MS   120
+#define TCP_WEDGED_DROP_MS     10000
+
+// True when lwIP can accept bytes for this socket right now. select() only reports
+// writable once at least TCP_SNDLOWAT (~2.8 KB) of send buffer is free, which holds a
+// whole companion frame (MAX_FRAME_SIZE + header), so a write issued after this probe
+// completes without blocking. Without the probe, WiFiClient::write() sits in 1 s
+// select() waits (up to 10) against a peer that stopped ACKing — a phone asleep in a
+// pocket keeps its socket "connected" but unwritable, and every mesh-RX push then
+// stalls the loop thread (and the UI) for seconds.
+static bool socketWritableNow(WiFiClient& client) {
+  int fd = client.fd();
+  if (fd < 0) return false;
+  fd_set wset;
+  FD_ZERO(&wset);
+  FD_SET(fd, &wset);
+  struct timeval tv;
+  tv.tv_sec = 0;
+  tv.tv_usec = 0;
+  return select(fd + 1, NULL, &wset, NULL, &tv) > 0 && FD_ISSET(fd, &wset);
+}
 
 static bool writeAllBytes(WiFiClient& client, const uint8_t* buf, size_t len, uint32_t timeout_ms) {
   size_t sent = 0;
   uint32_t start = millis();
   while (sent < len) {
     if (!client.connected()) return false;
-    size_t n = client.write(buf + sent, len - sent);
-    if (n > 0) {
-      sent += n;
-      continue;
+    if (socketWritableNow(client)) {
+      size_t n = client.write(buf + sent, len - sent);
+      if (n > 0) {
+        sent += n;
+        continue;
+      }
     }
     if (millis() - start >= timeout_ms) return false;
     delay(1);
@@ -26,6 +50,7 @@ TCPCompanionServer::TCPCompanionServer() : _server(WiFiServer()), _port(0), _pol
   for (int i = 0; i < TCP_COMPANION_MAX_CLIENTS; i++) {
     _clients[i].state = RECV_STATE_IDLE;
     _clients[i].in_use = false;
+    _clients[i].stall_ms = 0;
   }
 }
 
@@ -60,6 +85,7 @@ void TCPCompanionServer::acceptNewClients() {
       _clients[slot].in_use = true;
       _clients[slot].state = RECV_STATE_IDLE;
       _clients[slot].rx_len = 0;
+      _clients[slot].stall_ms = 0;
     } else {
       incoming.stop();
     }
@@ -81,6 +107,7 @@ void TCPCompanionServer::disconnectClient(int client_index) {
     _clients[client_index].client.stop();
     _clients[client_index].in_use = false;
     _clients[client_index].state = RECV_STATE_IDLE;
+    _clients[client_index].stall_ms = 0;
   }
 }
 
@@ -160,10 +187,21 @@ size_t TCPCompanionServer::writeToClient(int client_index, const uint8_t src[],
   WiFiClient* cl = &_clients[client_index].client;
   if (!writeAllBytes(*cl, hdr, 3, TCP_WRITE_TIMEOUT_MS) ||
       !writeAllBytes(*cl, src, len, TCP_WRITE_TIMEOUT_MS)) {
-    // Return 0 so caller can retry (e.g. contact list). Do not disconnect on transient
-    // buffer full; companion layer retries and will complete the sequence.
+    // Return 0 so caller can retry (e.g. contact list). A transient full buffer must
+    // not disconnect — the companion layer retries and completes the sequence. But a
+    // peer that stopped ACKing (phone asleep, walked out of range) keeps a half-open
+    // socket "connected" indefinitely while staying unwritable; drop it once the
+    // failure has been continuous for TCP_WEDGED_DROP_MS so it stops eating the
+    // write-timeout budget on every pushed frame.
+    TCPClientState* c = &_clients[client_index];
+    if (c->stall_ms == 0) {
+      c->stall_ms = millis() | 1;
+    } else if (millis() - c->stall_ms >= TCP_WEDGED_DROP_MS) {
+      disconnectClient(client_index);
+    }
     return 0;
   }
+  _clients[client_index].stall_ms = 0;
   return len;
 }
 

src/helpers/esp32/TCPCompanionServer.h

@@ -15,6 +15,7 @@ struct TCPClientState {
   uint16_t rx_len;
   uint8_t rx_buf[MAX_FRAME_SIZE];
   bool in_use;
+  uint32_t stall_ms;   // start of a continuous write-failure run; 0 = healthy
 };
 
 // Multi-client TCP server for companion protocol. Same frame format as serial.

src/helpers/esp32/WebSocketCompanionServer.cpp

@@ -2,12 +2,14 @@
 #include <Arduino.h>
 #include <mbedtls/sha1.h>
 #include <string.h>
+#include <lwip/sockets.h>
 
 #ifndef WS_FRAME_DEBUG
 #define WS_FRAME_DEBUG 0
 #endif
 
 #define TCP_WRITE_TIMEOUT_MS   120
+#define WS_WEDGED_DROP_MS      10000
 #define WS_MAGIC               "258EAFA5-E914-47DA-95CA-C5AB0DC85B11"
 
 // Browser opens http://device:8765 — short info page (plain WebSocket endpoint).
@@ -47,15 +49,32 @@ static void base64Encode20(const uint8_t* in, char* out) {
   *out = '\0';
 }
 
+// True when lwIP can accept bytes for this socket right now — see the identical probe
+// in TCPCompanionServer.cpp. Without it, WiFiClient::write() against a peer that
+// stopped ACKing (half-open socket) blocks in 1 s select() waits on the loop thread.
+static bool socketWritableNow(WiFiClient& client) {
+  int fd = client.fd();
+  if (fd < 0) return false;
+  fd_set wset;
+  FD_ZERO(&wset);
+  FD_SET(fd, &wset);
+  struct timeval tv;
+  tv.tv_sec = 0;
+  tv.tv_usec = 0;
+  return select(fd + 1, NULL, &wset, NULL, &tv) > 0 && FD_ISSET(fd, &wset);
+}
+
 static bool writeAllBytes(WiFiClient& client, const uint8_t* buf, size_t len, uint32_t timeout_ms) {
   size_t sent = 0;
   uint32_t start = millis();
   while (sent < len) {
     if (!client.connected()) return false;
-    size_t n = client.write(buf + sent, len - sent);
-    if (n > 0) {
-      sent += n;
-      continue;
+    if (socketWritableNow(client)) {
+      size_t n = client.write(buf + sent, len - sent);
+      if (n > 0) {
+        sent += n;
+        continue;
+      }
     }
     if (millis() - start >= timeout_ms) return false;
     delay(1);
@@ -72,6 +91,7 @@ WebSocketCompanionServer::WebSocketCompanionServer()
     _clients[i].handshake_len = 0;
     _clients[i].ws_state = WS_STATE_HEADER_0;
     _clients[i].comp_state = COMP_STATE_IDLE;
+    _clients[i].stall_ms = 0;
   }
 }
 
@@ -142,6 +162,7 @@ void WebSocketCompanionServer::acceptNewClients() {
     _clients[slot].handshake_len = 0;
     _clients[slot].ws_state = WS_STATE_HEADER_0;
     _clients[slot].comp_state = COMP_STATE_IDLE;
+    _clients[slot].stall_ms = 0;
   }
 }
 
@@ -355,8 +376,18 @@ size_t WebSocketCompanionServer::writeToClient(int client_index, const uint8_t s
 #if WS_FRAME_DEBUG
     Serial.printf("WS frame client=%d code=%u len=%u written=0\n", client_index, (unsigned)(len ? src[0] : 0), (unsigned)len);
 #endif
+    // Same wedged-peer reaper as the TCP server: a client that stays unwritable for
+    // WS_WEDGED_DROP_MS straight is half-open (browser tab gone, phone asleep) and
+    // would otherwise eat the write-timeout budget on every pushed frame forever.
+    WSClientState* c = &_clients[client_index];
+    if (c->stall_ms == 0) {
+      c->stall_ms = millis() | 1;
+    } else if (millis() - c->stall_ms >= WS_WEDGED_DROP_MS) {
+      disconnectClient(client_index);
+    }
     return 0;
   }
+  _clients[client_index].stall_ms = 0;
 #if WS_FRAME_DEBUG
   Serial.printf("WS frame client=%d code=%u len=%u written=%u\n", client_index, (unsigned)(len ? src[0] : 0), (unsigned)len, (unsigned)len);
 #endif
@@ -396,5 +427,6 @@ void WebSocketCompanionServer::disconnectClient(int client_index) {
   if (client_index >= 0 && client_index < WS_COMPANION_MAX_CLIENTS && _clients[client_index].in_use) {
     _clients[client_index].client.stop();
     _clients[client_index].in_use = false;
+    _clients[client_index].stall_ms = 0;
   }
 }

src/helpers/esp32/WebSocketCompanionServer.h

@@ -35,6 +35,7 @@ struct WSClientState {
   uint16_t comp_frame_len;
   uint16_t comp_rx_len;
   uint8_t comp_rx_buf[MAX_FRAME_SIZE];
+  uint32_t stall_ms;   // start of a continuous write-failure run; 0 = healthy
 };
 
 // WebSocket server for companion protocol. Same logical protocol as TCP; transport is RFC 6455 (plain WS).

src/main.cpp

@@ -573,9 +573,21 @@ void setup() {
 void loop() {
   // Run UI first every iteration so splash can dismiss at 3s even if mesh/serial blocks later (was stuck on version screen when the_mesh.loop() ran before ui_task.loop()).
 #ifdef DISPLAY_CLASS
-  ui_task.loop();
+  // ---- beta_31 field-freeze tracer (see UITask.cpp): time each loop section ----
+  extern void stallLog(const char* tag, uint32_t dur_ms);
+  extern const char* g_ui_stall_tag;
+  extern uint16_t    g_ui_stall_max;
+#define STALL_SCOPE(tag, call) { uint32_t _st0 = millis(); call; uint32_t _sdt = millis() - _st0; if (_sdt >= 200) stallLog(tag, _sdt); }
+  { uint32_t _ui0 = millis();
+    ui_task.loop();
+    uint32_t _uidt = millis() - _ui0;
+    if (_uidt >= 200) stallLog((g_ui_stall_max >= 150 && g_ui_stall_tag[0]) ? g_ui_stall_tag : "ui:other", _uidt);
+  }
 #endif
 #ifdef MULTI_TRANSPORT_COMPANION
+#ifdef DISPLAY_CLASS
+  uint32_t _wf0 = millis();   // beta_31 tracer: time the whole Wi-Fi state machine + SNTP span
+#endif
   static bool wifi_started = false;
   static uint32_t last_wifi_retry_ms = 0;
   static const uint32_t WIFI_RETRY_INTERVAL_MS = 10000;
@@ -700,14 +712,22 @@ void loop() {
       if (sntp_kicked && !sntp_pushed) sntp_kicked = false;
     }
   }
+#ifdef DISPLAY_CLASS
+  { uint32_t _wfdt = millis() - _wf0; if (_wfdt >= 200) stallLog("wifi-sm", _wfdt); }
+#endif
   // Defer TCP and WebSocket until after splash dismisses so the_mesh.loop() never blocks on accept() before ui_task.loop() runs.
   static const uint32_t TCP_DEFER_MS = 5000;   // 5 s: don't start TCP/WS until version screen has dismissed
   /* Only start TCP / WS when WiFi was actually brought up. In BLE-only mode
    * (no saved creds) the lwIP stack is never initialized — calling
    * WiFiServer::begin() crashes with a tcpip_adapter assert. */
   if (millis() > TCP_DEFER_MS && wifi_started) {
+#ifdef DISPLAY_CLASS
+    STALL_SCOPE("tcp-server", serial_interface.startTcpServer(WiFi.status() == WL_CONNECTED));
+    STALL_SCOPE("ws-tick",    serial_interface.tickWebSocketHandshake());
+#else
     serial_interface.startTcpServer(WiFi.status() == WL_CONNECTED);
     serial_interface.tickWebSocketHandshake();
+#endif
   }
 #endif
 #if defined(HAS_TOUCH_UI)
@@ -718,10 +738,18 @@ void loop() {
   // and clears the flag, so the next the_mesh.loop() re-arms RX correctly.
   if (!spectrumOwnsRadio())
 #endif
+#ifdef DISPLAY_CLASS
+  STALL_SCOPE("mesh", the_mesh.loop());
+#else
   the_mesh.loop();
+#endif
 #if defined(ESP32) && defined(MULTI_TRANSPORT_COMPANION)
+#ifdef DISPLAY_CLASS
+  STALL_SCOPE("mqtt", mqtt_bridge.loop());
+#else
   mqtt_bridge.loop();
 #endif
+#endif
 #if defined(GPS_BUF_DEBUG)
   // Bench diagnostic (build with -DGPS_BUF_DEBUG only; absent in releases): peak GPS UART
   // backlog accumulated between sensors.loop() drains. A peak above the old 256-byte default
@@ -738,7 +766,11 @@ void loop() {
     }
   }
 #endif
+#ifdef DISPLAY_CLASS
+  STALL_SCOPE("sensors", sensors.loop());
+#else
   sensors.loop();
+#endif
 #if defined(ESP32)
   // GPS time guard (Ricky Leong's "stuck at 1902"): MicroNMEALocationProvider
   // sets the mesh RTC from a GPS *position* fix even before the date fields are
@@ -795,6 +827,10 @@ void loop() {
   // Idle light-sleep gate: evaluated every loop tick. g_enabled is false by
   // default (Task 1 is inert); Task 2 sets it from the NVS pref and wires
   // the real predicates so the gate can actually pass and arm light sleep.
+#ifdef DISPLAY_CLASS
+  STALL_SCOPE("sleep-gate", touchSleep::loopEnd(millis()));
+#else
   touchSleep::loopEnd(millis());
 #endif
+#endif
 }

src/ui-touch/UITask.cpp

@@ -8553,6 +8553,11 @@ static void openMemoryDetailCb(lv_event_t* e) {
 // Render the System-info text into buf. Re-callable so live values (uptime,
 // free heap) update; the inline About page calls it once at build then
 // refreshSysInfo() re-runs it ~1 Hz while that tab is visible.
+// beta_31 field-freeze tracer record (ring + logger defined with UITask::loop below).
+struct StallRec { uint32_t at_s; uint16_t dur_ms; const char* tag; };
+extern StallRec g_stall_ring[16];
+extern uint8_t  g_stall_cnt, g_stall_w;
+
 static void sysInfoText(char* buf, size_t cap) {
   int p = 0;
 #if defined(ESP32)
@@ -8644,6 +8649,21 @@ static void sysInfoText(char* buf, size_t cap) {
   p += snprintf(buf + p, cap - p,
                 "Last reset\n  %s\n\n", resetReasonString(esp_reset_reason()));
 #endif
+  // beta_31 field-freeze tracer: recent loop stalls (>0.2 s), newest first — a
+  // field tester photographs this instead of needing a serial console.
+  p += snprintf(buf + p, cap - p, "Loop stalls (>0.2s)\n");
+  if (g_stall_cnt == 0) {
+    p += snprintf(buf + p, cap - p, "  none recorded\n\n");
+  } else {
+    const uint32_t nows = millis() / 1000u;
+    for (int i = 0; i < (int)g_stall_cnt && i < 6; ++i) {
+      const int idx = ((int)g_stall_w - 1 - i + 32) % 16;
+      const StallRec& r = g_stall_ring[idx];
+      p += snprintf(buf + p, cap - p, "  -%lus  %s  %ums\n",
+                    (unsigned long)(nows - r.at_s), r.tag, (unsigned)r.dur_ms);
+    }
+    p += snprintf(buf + p, cap - p, "\n");
+  }
   p += snprintf(buf + p, cap - p,
                 "Build\n  %s\n  %s\n  WADAMESH TOUCH\n",
                 FIRMWARE_VERSION, FIRMWARE_BUILD_DATE);
@@ -33818,16 +33838,20 @@ void UITask::markThreadsDirty(unsigned long delay_ms) {
 }
 
 void UITask::markMsgsDirty(unsigned long delay_ms) {
-  // Deep (SD) ring: a flush rewrites the whole ring file (~1.1 MB at 5000 slots),
-  // so space the writes out — at least 30 s apart instead of the 0.5-2 s cadence
-  // of the small internal-flash ring. persistHistoryNow() still forces a write on
-  // reboot/shutdown, so at most the last <30 s of messages ride on a hard crash.
-  if (_ui_msg_cap > MAX_UI_MESSAGES && delay_ms < 30000) delay_ms = 30000;
+  // Every flush rewrites the WHOLE ring file on the loop thread, so space the
+  // writes out: at least 30 s for the deep SD ring (~1.1 MB at 5000 slots) and at
+  // least 10 s for the small internal-flash ring (~115 KB) — a busy channel used
+  // to trigger a full rewrite (+ possible SPIFFS GC) ~2 s after EVERY message,
+  // which read as "the device freezes 1-2 s every ~10 s" on stable beta_31.
+  // persistHistoryNow() still forces a write on reboot/shutdown, so a hard crash
+  // costs at most the last few seconds of messages.
+  if (_ui_msg_cap > MAX_UI_MESSAGES) { if (delay_ms < 30000) delay_ms = 30000; }
+  else                               { if (delay_ms < 10000) delay_ms = 10000; }
   const unsigned long deadline = millis() + delay_ms;
   if (!_msgs_dirty || deadline < _next_msgs_flush_ms)
     _next_msgs_flush_ms = deadline;
   _msgs_dirty = true;
-  markThreadsDirty(200);  // thread state (last_ts, unread) also changed
+  markThreadsDirty(1500);  // thread state (last_ts, unread) changed too — coalesce a message burst into one write
 }
 
 void UITask::flushHistoryIfDue(unsigned long now) {
@@ -34203,10 +34227,14 @@ bool UITask::saveThreadsToStorage() {
 
 bool UITask::saveMsgsToStorage() {
 #if defined(ESP32)
-  // Write from internal-RAM stack structs (fast). NOTE: do NOT "optimize"
-  // this into a single big write from a PSRAM buffer — the flash driver has
-  // to bounce a PSRAM source through internal RAM chunk-by-chunk, which is
-  // SLOWER than these per-struct writes and was a real regression.
+  // Chunked writer (same pattern that fixed the contacts stall, #82): records are
+  // packed into an INTERNAL-RAM chunk and flushed in ~6 KB writes, so the ring
+  // costs ~20 FS calls (500 slots) / ~210 (5000 slots) instead of one small write
+  // per record — a full-ring rewrite froze the UI 1-2 s per incoming message on
+  // busy meshes. This is NOT the "one big write from PSRAM" that regressed before:
+  // the chunk lives in internal RAM (the flash driver never bounces a PSRAM
+  // source), and the per-record field copies are unchanged. Alloc failure falls
+  // back to the original per-record writes.
   WdtHeavyGuard _wg;   // a fragmenting write can trigger a multi-second SPIFFS GC
   File f = uiDataOpen(k_ui_msgs_path, "w");
   if (!f) return false;
@@ -34222,28 +34250,59 @@ bool UITask::saveMsgsToStorage() {
     f.close(); return false;
   }
 
-  UiHistoryMsg m{};
-  for (int i = 0; i < _ui_msg_cap; ++i) {
-    memset(&m, 0, sizeof(m));
-    m.ts          = _ui_msgs[i].ts;
-    m.channel     = _ui_msgs[i].channel ? 1u : 0u;
-    m.outgoing    = _ui_msgs[i].outgoing ? 1u : 0u;
-    m.meta_flags  = _ui_msgs[i].meta_flags;
-    m.path_len    = _ui_msgs[i].path_len;
-    m.snr_q4      = _ui_msgs[i].snr_q4;
-    m.rssi        = _ui_msgs[i].rssi;
-    strncpy(m.thread, _ui_msgs[i].thread, MAX_THREAD_NAME);
-    m.thread[MAX_THREAD_NAME] = '\0';
-    strncpy(m.sender, _ui_msgs[i].sender, MAX_SENDER_NAME);
-    m.sender[MAX_SENDER_NAME] = '\0';
-    strncpy(m.text, _ui_msgs[i].text, MAX_MSG_TEXT);
-    m.text[MAX_MSG_TEXT] = '\0';
-    if (f.write(reinterpret_cast<const uint8_t*>(&m), sizeof(m)) != sizeof(m)) {
-      f.close(); return false;
+  const size_t REC = sizeof(UiHistoryMsg);
+  size_t chunk_recs = 6144 / REC;
+  if (chunk_recs < 1) chunk_recs = 1;
+  uint8_t* buf = (uint8_t*)malloc(REC * chunk_recs);   // internal RAM by default
+  bool ok = true;
+  if (buf) {
+    size_t fill = 0;
+    for (int i = 0; ok && i < _ui_msg_cap; ++i) {
+      UiHistoryMsg* m = reinterpret_cast<UiHistoryMsg*>(buf + fill);
+      memset(m, 0, REC);
+      m->ts          = _ui_msgs[i].ts;
+      m->channel     = _ui_msgs[i].channel ? 1u : 0u;
+      m->outgoing    = _ui_msgs[i].outgoing ? 1u : 0u;
+      m->meta_flags  = _ui_msgs[i].meta_flags;
+      m->path_len    = _ui_msgs[i].path_len;
+      m->snr_q4      = _ui_msgs[i].snr_q4;
+      m->rssi        = _ui_msgs[i].rssi;
+      strncpy(m->thread, _ui_msgs[i].thread, MAX_THREAD_NAME);
+      m->thread[MAX_THREAD_NAME] = '\0';
+      strncpy(m->sender, _ui_msgs[i].sender, MAX_SENDER_NAME);
+      m->sender[MAX_SENDER_NAME] = '\0';
+      strncpy(m->text, _ui_msgs[i].text, MAX_MSG_TEXT);
+      m->text[MAX_MSG_TEXT] = '\0';
+      fill += REC;
+      if (fill == REC * chunk_recs) {
+        ok = (f.write(buf, fill) == fill);
+        fill = 0;
+      }
+    }
+    if (ok && fill > 0) ok = (f.write(buf, fill) == fill);
+    free(buf);
+  } else {
+    UiHistoryMsg m{};
+    for (int i = 0; ok && i < _ui_msg_cap; ++i) {
+      memset(&m, 0, sizeof(m));
+      m.ts          = _ui_msgs[i].ts;
+      m.channel     = _ui_msgs[i].channel ? 1u : 0u;
+      m.outgoing    = _ui_msgs[i].outgoing ? 1u : 0u;
+      m.meta_flags  = _ui_msgs[i].meta_flags;
+      m.path_len    = _ui_msgs[i].path_len;
+      m.snr_q4      = _ui_msgs[i].snr_q4;
+      m.rssi        = _ui_msgs[i].rssi;
+      strncpy(m.thread, _ui_msgs[i].thread, MAX_THREAD_NAME);
+      m.thread[MAX_THREAD_NAME] = '\0';
+      strncpy(m.sender, _ui_msgs[i].sender, MAX_SENDER_NAME);
+      m.sender[MAX_SENDER_NAME] = '\0';
+      strncpy(m.text, _ui_msgs[i].text, MAX_MSG_TEXT);
+      m.text[MAX_MSG_TEXT] = '\0';
+      ok = (f.write(reinterpret_cast<const uint8_t*>(&m), sizeof(m)) == sizeof(m));
     }
   }
   f.close();
-  return true;
+  return ok;
 #else
   return false;
 #endif
@@ -36573,8 +36632,35 @@ void UITask::notify(UIEventType t) {
 // ============================================================
 // UITask::loop
 // ============================================================
+// ---- beta_31 field-freeze tracer (diagnostics only, no behaviour change) ----
+// Any instrumented main-loop section that blocks >200 ms lands in this ring,
+// printed to Serial as [STALL] and listed at the bottom of Settings -> About so
+// a field tester can simply photograph the screen. uiCp() checkpoints attribute
+// a slow UITask::loop pass to the section that actually ate the time.
+StallRec g_stall_ring[16];   // struct defined above sysInfoText (About shows the ring)
+uint8_t  g_stall_cnt = 0, g_stall_w = 0;
+void stallLog(const char* tag, uint32_t dur_ms) {
+  if (dur_ms < 200) return;
+  g_stall_ring[g_stall_w] = { (uint32_t)(millis() / 1000u), (uint16_t)(dur_ms > 65535 ? 65535 : dur_ms), tag };
+  g_stall_w = (uint8_t)((g_stall_w + 1) % 16);
+  if (g_stall_cnt < 16) g_stall_cnt++;
+  Serial.printf("[STALL] %s %lums\n", tag, (unsigned long)dur_ms);
+}
+const char*   g_ui_stall_tag = "";     // slowest section of the current UITask::loop pass
+uint16_t      g_ui_stall_max = 0;
+static const char*  s_ui_cp_tag = "";
+static uint32_t     s_ui_cp_t0  = 0;
+static inline void uiCp(const char* next) {
+  const uint32_t nowms = millis();
+  const uint32_t dt = nowms - s_ui_cp_t0;
+  if (dt > g_ui_stall_max) { g_ui_stall_max = (uint16_t)(dt > 65535 ? 65535 : dt); g_ui_stall_tag = s_ui_cp_tag; }
+  s_ui_cp_tag = next;
+  s_ui_cp_t0  = nowms;
+}
+
 void UITask::loop() {
   unsigned long now = millis();
+  g_ui_stall_max = 0; g_ui_stall_tag = ""; s_ui_cp_tag = "ui:head"; s_ui_cp_t0 = now;
 #if defined(ESP32)
   // Persist the clock floor every 15 min (small prefs-blob write; SdNvsPrefs
   // rewrites a file on the Launcher-SD backend, so keep the cadence low). Power
@@ -36600,7 +36686,9 @@ void UITask::loop() {
   // but if the page ever vanished without that, the mesh would stay off the
   // radio forever — so re-apply the mesh params and release ownership here.
   if (s_spectrum_active && !s_spec_root) spectrumRestoreRadio();
+  uiCp("ui:hist");
   flushHistoryIfDue(now);
+  uiCp("ui:ctdirty");
   // A contact was discovered/updated via an advert — possibly while a companion app was connected
   // over BLE (issue #73). FORCE the rebuild: an advert can fill in a name, or refresh an existing
   // (or app-added) contact, WITHOUT changing getNumContacts() — which the count-cached
@@ -36634,7 +36722,9 @@ void UITask::loop() {
   if (s_msgled_flash_until) msgLedRefresh(getUnreadTotal() > 0);   // end the one-shot envelope-LED flash on time
 #endif
   { static bool s_disc_loaded = false; if (!s_disc_loaded) { s_disc_loaded = true; loadDiscovered(); } }
+  uiCp("ui:disc");
   discoveredFlushIfDue(now);   // persist the discovered ring (rate-capped) so it survives reboot
+  uiCp("ui:gps");
   updateGpsLocation(now);   // sync + persist node location from GPS once fixed
   ++s_live_diag_loops;
   if (_alert_expiry != 0 && now >= _alert_expiry) {
@@ -37100,7 +37190,9 @@ void UITask::loop() {
 #if CAP_SD
   telemetryPollTick((uint32_t)now); // auto-poll due nodes -> log (no window)
 #endif
+  uiCp("ui:verchk");
   versionCheckService(now);   // firmware update check (gear badge + About line)
+  uiCp("ui:sbar");
   refreshSysInfo(now);        // live uptime / heap on the About sub-tab
 #if defined(HAS_TDECK_GT911)
   refreshSleepDiag(now);     // live sleep counters on the Lock settings panel
@@ -37177,7 +37269,9 @@ void UITask::loop() {
   // (fed by handleHwKey) or the trackball D-pad nav (fed by updateTrackball -> navMoveDir).
   if (s_kbd_nav || s_tb_nav) navMaybeRebuild();
 #endif
+  uiCp("ui:lvgl");
   lv_timer_handler();
+  uiCp("ui:tail");
 #if defined(HAS_TANMATSU)
   if (s_nav_entered_obj) {
     lv_obj_t* ent = s_nav_entered_obj; s_nav_entered_obj = nullptr;