fix(tracking): skip bad Firefox carets and container-sized focus (#278)

docs/TRACKING-FINDINGS.md

@@ -48,3 +48,17 @@ highlighted option through UIA, so Wind follows only the terminal caret there.
   cursor's visible body (opaque bounds around the hotspot, re-measured on cursor change).
 - **Margin.** Edge mode has its own `mouseMarginPct` (default 0: the cursor reaches the view edge
   before the view moves; Settings slider 0-30%). `trackMarginPct` (15%) stays the caret/focus one.
+
+## Firefox in a zoomed iframe (issue #278, 2026-09-29)
+
+- A Claude artifact (an iframe) at high Ctrl+ page zoom in Zen: Firefox reports the caret wrongly
+  from BOTH sources. The Win32 caret sits below and right of the input (input 2226,997 798x74,
+  caret 3097,1226 1x118); the UIA caret range and the element bounds point above the real text.
+  Not reproducible on a plain page at the same zoom. There is no correct source to fall back to.
+- Fix: in Gecko windows (`MozillaWindowClass`, Firefox and all forks) a caret whose centre is outside
+  its own element is skipped, so the view stays put. Limited to Gecko so no app that tracked
+  correctly before can lose tracking. A UIA fallback was tried and field-rejected (it lands above).
+- Leaving a text box moves focus to the whole page (3400x1912); centring on it dropped the view.
+  Focus rects covering half the monitor or more are containers and are skipped (all apps).
+- Diagnostic tool from this hunt: a recorder logging the foreground process, Win32 caret, UIA caret
+  and focus bounds every 100 ms; the Win32/UIA disagreement is what located the bad source.

src/focus_track.cpp

@@ -1,5 +1,6 @@
 #include "focus_track.h"
 #include "logging.h"
+#include "track_filter.h"
 #include <windows.h>
 #include <objbase.h>
 #include <oleacc.h>
@@ -45,6 +46,12 @@ static bool IsOwnOrTooltip(HWND h) {
     return wcscmp(cls, L"tooltips_class32") == 0 || wcscmp(cls, L"Xaml_WindowedPopupClass") == 0;
 }
 
+// A Gecko browser window (Firefox and every fork share this class).
+static bool IsGeckoWindow(HWND h) {
+    wchar_t cls[64] = {}; GetClassNameW(h, cls, 64);
+    return wcscmp(cls, L"MozillaWindowClass") == 0;
+}
+
 // Classic Win32 caret of the foreground thread, in screen px. False when there is none.
 static bool Win32Caret(RECT& out) {
     HWND fg = GetForegroundWindow();
@@ -181,11 +188,21 @@ void FocusTracker::run() {
         if (IsOwnOrTooltip(fg)) return;
         IUIAutomationElement* el = nullptr;
         if (uia) uia->GetFocusedElement(&el);
+        RECT b{};                                   // the focused element's bounds (empty = unknown)
+        if (el && FAILED(el->get_CurrentBoundingRectangle(&b))) b = RECT{};
         // 1. A focus change: follow the focused control (if wanted). Its caret becomes the baseline.
-        if (focusChanged && wantFocus_.load() && el) {
-            RECT b{};
-            if (SUCCEEDED(el->get_CurrentBoundingRectangle(&b)) && b.right > b.left && b.bottom > b.top)
+        //    A container-sized focus (the page after leaving a text box, a pane, the window) is not
+        //    something the user moved to, so it is skipped (issue #278).
+        if (focusChanged && wantFocus_.load() && el && b.right > b.left && b.bottom > b.top) {
+            MONITORINFO mi{ sizeof(mi) };
+            const bool haveMon = GetMonitorInfoW(MonitorFromWindow(fg, MONITOR_DEFAULTTONEAREST), &mi) != 0;
+            if (haveMon && IsContainerFocus({ b.left, b.top, b.right, b.bottom },
+                                            mi.rcMonitor.left, mi.rcMonitor.top, mi.rcMonitor.right, mi.rcMonitor.bottom)) {
+                if (log_.load()) wind::Log(wind::LogLevel::Info, "track", "focus skipped (container): %ld,%ld %ldx%ld",
+                                           b.left, b.top, b.right - b.left, b.bottom - b.top);
+            } else {
                 publish(TrackKind::Focus, b.left, b.top, b.right, b.bottom, "uia-focus");
+            }
         }
         // 2. The caret: published only when it moved within the same focus.
         if (wantCaret_.load()) {
@@ -196,7 +213,18 @@ void FocusTracker::run() {
                     if (log_.load()) wind::Log(wind::LogLevel::Info, "track", "caret baseline via %s: %ld,%ld", src, rc.left, rc.top);
                 } else if (!EqualRect(&rc, &lastCaret)) {
                     lastCaret = rc;
-                    publish(TrackKind::Caret, rc.left, rc.top, rc.right, rc.bottom, src);
+                    // In a Gecko browser (Firefox, Zen, LibreWolf...: one window class), a caret
+                    // outside its own element is a bad report, not a place to look (a zoomed iframe,
+                    // issue #278; its UIA caret was wrong too, so there is no source to fall back
+                    // to): the view stays where it is. Other apps are unchanged, so no case that
+                    // worked before can lose tracking to this rule.
+                    if (IsGeckoWindow(fg) &&
+                        !CaretInsideElement({ rc.left, rc.top, rc.right, rc.bottom }, { b.left, b.top, b.right, b.bottom })) {
+                        if (log_.load()) wind::Log(wind::LogLevel::Info, "track", "caret skipped (outside its element %ld,%ld %ldx%ld) via %s: %ld,%ld",
+                                                   b.left, b.top, b.right - b.left, b.bottom - b.top, src, rc.left, rc.top);
+                    } else {
+                        publish(TrackKind::Caret, rc.left, rc.top, rc.right, rc.bottom, src);
+                    }
                 }
             }
         }

src/track_filter.h

@@ -0,0 +1,30 @@
+#pragma once
+// Sanity filters for what the focus/caret watcher reports (issue #278). Pure; tests/test_track_filter.cpp.
+namespace wind {
+
+struct TrackBox { long l, t, r, b; };
+
+// Is the caret where its own focused element is? Firefox inside a zoomed iframe (a Claude artifact at
+// high Ctrl+ page zoom, field 2026-09-29) reports the caret through BOTH GetGUIThreadInfo and UIA well
+// outside the focused input (input 2226,997 798x74, caret 3097,1226 1x118), so following it showed
+// empty page. A caret whose centre lies outside the element (plus a small slack for borders and
+// padding) is not trusted. An empty element rect (unknown) trusts the caret.
+inline bool CaretInsideElement(const TrackBox& caret, const TrackBox& elem, long slack = 8) {
+    if (elem.r <= elem.l || elem.b <= elem.t) return true;
+    const long cx = (caret.l + caret.r) / 2, cy = (caret.t + caret.b) / 2;
+    return cx >= elem.l - slack && cx <= elem.r + slack && cy >= elem.t - slack && cy <= elem.b + slack;
+}
+
+// Is a focus rect a CONTAINER (the page, a pane, the window) rather than a control someone moved to?
+// Leaving a text area moves focus to the whole document, and centring on it dropped the view to the
+// middle of the page (field 2026-09-29: 3400x1912 on a 3840x2160 monitor). Half the monitor's area,
+// counting only the part on the monitor, is the line.
+inline bool IsContainerFocus(const TrackBox& rc, long monL, long monT, long monR, long monB) {
+    const long l = rc.l > monL ? rc.l : monL, r = rc.r < monR ? rc.r : monR;
+    const long t = rc.t > monT ? rc.t : monT, b = rc.b < monB ? rc.b : monB;
+    if (r <= l || b <= t) return false;
+    const double area = (double)(r - l) * (double)(b - t);
+    const double mon = (double)(monR - monL) * (double)(monB - monT);
+    return mon > 0 && area >= 0.5 * mon;
+}
+}  // namespace wind

src/version.h

@@ -4,7 +4,7 @@
 
 #define WIND_VER_MAJOR 0
 #define WIND_VER_MINOR 10
-#define WIND_VER_PATCH 3
+#define WIND_VER_PATCH 4
 
 // String form for logs/snapshot/UI. Keep in sync with the numeric parts above.
-#define WIND_VERSION_STR "0.10.3"
+#define WIND_VERSION_STR "0.10.4"

tests/test_track_filter.cpp

@@ -0,0 +1,20 @@
+#include "doctest.h"
+#include "../src/track_filter.h"
+using namespace wind;
+
+TEST_CASE("caret inside its element is trusted; the field-recorded Firefox iframe caret is not") {
+    const TrackBox input{ 2226, 997, 3024, 1071 };
+    CHECK(CaretInsideElement({ 2453, 991, 2478, 1077 }, input));            // the good reading
+    CHECK_FALSE(CaretInsideElement({ 3097, 1226, 3098, 1344 }, input));     // below and right of it
+    CHECK(CaretInsideElement({ 3020, 1000, 3022, 1068 }, input));           // at the right end, inside
+    CHECK(CaretInsideElement({ 5, 5, 6, 20 }, { 0, 0, 0, 0 }));             // unknown element: trust
+}
+TEST_CASE("terminal caret beside its cell stays trusted") {
+    CHECK(CaretInsideElement({ 332, 1692, 343, 1737 }, { 322, 1692, 346, 1737 }));
+}
+TEST_CASE("page-sized focus is a container; controls and huge zoomed inputs partly off-screen are not") {
+    CHECK(IsContainerFocus({ 422, 122, 3822, 2034 }, 0, 0, 3840, 2160));      // the document
+    CHECK_FALSE(IsContainerFocus({ 2226, 997, 3024, 1071 }, 0, 0, 3840, 2160));
+    CHECK_FALSE(IsContainerFocus({ 902, 912, 6146, 2106 }, 0, 0, 3840, 2160)); // 5244x1194, 41% on screen
+    CHECK_FALSE(IsContainerFocus({ 5000, 0, 6000, 100 }, 0, 0, 3840, 2160));  // off the monitor
+}