diff --git a/rules/gui/gl-teardown-must-run-on-the-refresh-thread.md b/rules/gui/gl-teardown-must-run-on-the-refresh-thread.md
new file mode 100644
index 0000000..4ff10d7
--- /dev/null
+++ b/rules/gui/gl-teardown-must-run-on-the-refresh-thread.md
@@ -0,0 +1,54 @@
+# GL teardown must run on the REX3-Refresh thread
+
+**Keywords:** opengl,gl,glow,glutin,glDeleteTextures,thread,context,make_current,refresh,poweroff,shutdown,crash,segfault,compositor,destroy,rex3,renderer
+**Category:** gui
+
+## All OpenGL calls — including teardown — belong to the thread that owns the context
+
+The GL context is created and made current lazily by `GlRenderer::init_gl`, which
+runs from `present()` — and `present()` is only ever called from `Rex3::refresh_loop`
+(the **REX3-Refresh** thread). That thread is the sole owner of the context; this is
+documented by the `unsafe impl Send for GlRenderer` comment ("sent to the refresh
+thread where it owns and uses the GL context. No other thread touches these fields").
+
+On macOS, GL functions dispatch through the **calling thread's** current context.
+A GL call from a thread with no current context dereferences a null current-context
+struct and faults at a small offset (the observed crash was `glDeleteTextures` with
+`KERN_INVALID_ADDRESS at 0x1e0` — a null + field offset, not a freed pointer).
+
+## The bug this rule encodes
+
+`Rex3::stop()` used to:
+1. set `running = false`, join the REX3-Processor thread,
+2. **join the REX3-Refresh thread** (the context owner is now gone), then
+3. call `renderer.stop()` → `Compositor::destroy(&state.gl)` → `glDeleteTextures`
+   **on whatever thread called `Rex3::stop()`** — never the refresh thread.
+
+`Rex3::stop()` is reached from many non-owning threads: the `machine-events` thread
+(guest soft power-off → `MachineEvent::PowerOff` → `Machine::stop`), the **main**
+thread (`main.rs` after the winit loop returns on window close), and the monitor/CI
+thread (`reset`, snapshot load/save). The GUI power-off and window-close paths have a
+live context and crash; the headless/CI paths have `state == None` so teardown is a
+silent no-op, which is why it went unnoticed.
+
+## The fix
+
+Tear the renderer down at the **end of `refresh_loop`**, right after the
+`while self.running` loop exits — i.e. on the owning thread, while the context is
+still current — and drop the `renderer.stop()` call from `Rex3::stop()`. Ordering is
+preserved because `Rex3::stop()` joins the refresh thread, so teardown still finishes
+before `stop()` returns. After a `reset`, `restart_peripherals()` calls
+`rex3.start()`, which respawns the refresh thread; `present()` then re-inits GL lazily.
+
+## Still open (same root cause, different trigger)
+
+`disp compositor <gl|sw>` → `GlRenderer::switch_compositor` (ui.rs) calls
+`compositor.destroy(&state.gl)` directly on the **monitor** thread. Same foreign-thread
+GL teardown, same fault; rarely triggered. A proper fix routes the compositor swap
+through the refresh thread (e.g. a pending-request flag the loop services).
+
+## Rule of thumb
+
+Never call any `gl.*` / `glow` method (create, upload, draw, or delete) from outside
+`refresh_loop`/`present`. If another thread needs GL work done, signal the refresh
+thread and let it do it.
diff --git a/src/rex3.rs b/src/rex3.rs
index 4023c0a..646daab 100644
--- a/src/rex3.rs
+++ b/src/rex3.rs
@@ -3625,6 +3625,10 @@ impl Rex3 {
                 thread::sleep(frame_duration - elapsed);
             }
         }
+
+        if let Some(renderer) = self.renderer.lock().as_mut() {
+            renderer.stop();
+        }
     }
 
     /// Returns `true` if the register offset was recognized and updated state.
@@ -3897,10 +3901,6 @@ impl Device for Rex3 {
         #[cfg(feature = "developer")]
         eprintln!("REX3: GFIFO high-watermark = {} / {} entries",
             self.gfifo_hwm.load(Ordering::Relaxed), GFIFO_DEPTH);
-
-        if let Some(renderer) = self.renderer.lock().as_mut() {
-            renderer.stop();
-        }
     }
 
     fn start(&self) {