diff --git a/frontend/src/components/editor/EditorToolbar.tsx b/frontend/src/components/editor/EditorToolbar.tsx
index e9ece5c6..429860fd 100644
--- a/frontend/src/components/editor/EditorToolbar.tsx
+++ b/frontend/src/components/editor/EditorToolbar.tsx
@@ -811,6 +811,19 @@ export const EditorToolbar = ({
       // QEMU boards: auto-compile if no firmware available yet
       if (isQemuBoard) {
         console.log('[handleRun] QEMU path');
+        // Clean restart when the board is already running. Esp32Bridge.connect()
+        // is a no-op while the socket is non-CLOSED, so startBoard() on a live
+        // session does NOTHING — and if the backend QEMU session has since died
+        // but the frontend socket is still zombie (CONNECTING/OPEN/CLOSING), the
+        // user sees a dead sim that only a page reload fixes. This is the exact
+        // "el agente terminó, di Run y no funcionó; recargué y sí" report: the
+        // agent's run_simulation left the board running, so the user's Run
+        // no-op'd. Stop first (closes the WS), let it settle, then boot fresh —
+        // mirrors what the MicroPython branch above already does.
+        if (board?.running) {
+          stopBoard(activeBoardId);
+          await new Promise((resolve) => setTimeout(resolve, 300));
+        }
         if (!board?.compiledProgram || codeChangedSinceLastCompile) {
           console.log('[handleRun] auto-compile + run');
           autoRunAfterCompile.current = true;
diff --git a/frontend/src/components/simulator/SimulatorCanvas.tsx b/frontend/src/components/simulator/SimulatorCanvas.tsx
index 414b50b7..8fc86ed2 100644
--- a/frontend/src/components/simulator/SimulatorCanvas.tsx
+++ b/frontend/src/components/simulator/SimulatorCanvas.tsx
@@ -83,6 +83,16 @@ import './SimulatorCanvas.css';
 /** World-units of tolerance for alignment snap (scales with zoom). */
 const ALIGN_SNAP_PX = 6;
 
+/**
+ * Large-bodied display parts whose face should occlude wires when the part is
+ * seated on a breadboard — the flat canvas has no depth, so bridge wires
+ * routed to holes UNDER the body would otherwise paint over the readout.
+ * Matches both the metadata ids (`7segment`, `ssd1306`, …) and the `wokwi-`/
+ * `velxio-` element-name variants. Thin two-pin parts are deliberately absent.
+ */
+const DISPLAY_BODY_METADATA_RE =
+  /(7segment|led-matrix|max7219|neopixel-matrix|ssd1306|oled|ili9341|lcd1602|lcd2004|led-ring)/i;
+
 /** Long-press duration for touch context menu (ms). */
 const LONG_PRESS_MS = 500;
 
@@ -2240,7 +2250,26 @@ export const SimulatorCanvas = ({ headerSlot }: SimulatorCanvasProps = {}) => {
     // them — so they always sit at the very back: below boards (z 0), other
     // components (z 1/2) and wires (z 35). Selection doesn't raise them.
     const isBreadboard = String(component.metadataId).startsWith('breadboard');
-    const groupZIndex = isBreadboard ? -1 : isSelected ? 2 : 1;
+    // A large-bodied display seated on a breadboard physically sits ON the
+    // board, so wires routed to holes UNDERNEATH it pass behind its body in
+    // real life. On the flat canvas those bridge wires (segment strips whose
+    // holes are literally under the display) otherwise paint OVER the digits
+    // — the reported "los cables se ven superpuestos y casi ni se ven los
+    // dígitos". Raise a seated display above the wire layer (z 35) so its
+    // face occludes the wires crossing it, exactly as the real part would.
+    // Scoped to occluding display bodies + only when actually seated, so
+    // thin parts (resistors, LEDs) and free-floating displays are untouched.
+    const isSeated = (seatedPinsByComponent.get(component.id)?.length ?? 0) > 0;
+    const occludesWhenSeated = DISPLAY_BODY_METADATA_RE.test(String(component.metadataId));
+    const groupZIndex = isBreadboard
+      ? -1
+      : isSeated && occludesWhenSeated
+        ? isSelected
+          ? 37
+          : 36
+        : isSelected
+          ? 2
+          : 1;
 
     return (
       <React.Fragment key={component.id}>