fix: don't start the app from main() on android; use the real inset retry

main() must not run real_main on android. The library's DT_INIT points at Odin's
_odin_entry_point, which the loader runs the moment NativeActivity dlopens
libmain.so -- before the glue calls android_main, so context.user_ptr is nil and
backend_init faults writing app.onAppCmd. The proc still has to exist: package
main requires it and it is the desktop entry point.

The insets retry now uses request_render_at. A plain request_render made before
the build is consumed by the frame it is running in, so it schedules nothing and
the loop parks with insets unresolved; it only appeared to work because the
first frames redraw anyway while the tree settles.

Guards the JNI-only paths (show_toast, refresh_insets) behind the android
subtarget so the shared real_main is usable on desktop. refresh_insets returns
true there rather than false: zero insets are correct where there are no system
bars, and returning false would retry every frame forever.

Also declares appCategory=game, and enables the engine's frame timings.

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

android/AndroidManifest.xml

@@ -13,10 +13,11 @@
 
   <application
       android:label="Dogfooding"
+      android:appCategory="game"
       android:hasCode="false"
       android:icon="@mipmap/ic_launcher"
       android:roundIcon="@mipmap/ic_launcher_round"
-      android:debuggable="false">
+      android:debuggable="true">
 
     <activity android:name="android.app.NativeActivity"
               android:configChanges="orientation|keyboardHidden"

main.odin

@@ -2,6 +2,7 @@ package main
 
 import "shared:android"
 import "base:runtime"
+import "core:time"
 import engine "shared:zephr"
 import "jni"
 
@@ -18,7 +19,18 @@ insets: Edge_Insets
 insets_valid: bool
 
 main :: proc() {
+    // Android must not start the app from here. This library's DT_INIT points at
+    // Odin's _odin_entry_point, which the dynamic loader runs the moment
+    // NativeActivity dlopens libmain.so -- long before the glue spawns its
+    // thread and calls android_main, and so before context.user_ptr holds the
+    // android_app. Running real_main here reaches backend_init with a nil app
+    // and faults writing app.onAppCmd.
+    //
+    // The proc still has to exist: package main requires it, and it is the
+    // desktop entry point.
+    when ODIN_PLATFORM_SUBTARGET != .Android {
         real_main()
+    }
 }
 
 real_main :: proc() {
@@ -41,6 +53,11 @@ real_main :: proc() {
     // flat out forever, which on a phone is straight battery burn.
     engine.set_render_on_change(true)
 
+    // Rolling work-vs-swap timings to logcat, to tell a slow frame apart from a
+    // frame that is merely waiting on vsync.
+    engine.set_frame_stats(true)
+
+
     for !engine.should_quit() {
         engine.frame_start()
         defer engine.frame_end()
@@ -60,11 +77,16 @@ real_main :: proc() {
         if !insets_valid {
             insets_valid = refresh_insets()
             if !insets_valid {
-                // The decor view attaches asynchronously and produces no event
-                // when it does. Under render_on_change the loop would park
-                // before the retry ever ran, leaving the app bar with no status
-                // bar padding.
-                engine.request_render()
+                // The decor view attaches asynchronously and emits no event when
+                // it does, so the retry has to drive itself.
+                //
+                // Deliberately request_render_at, not request_render: a plain
+                // request made here -- before the build -- is consumed by the
+                // very frame it is running in, so it schedules nothing and the
+                // loop parks with insets still unresolved. A timed wake is the
+                // one that survives into the next frame. It only works today
+                // because the first frames redraw anyway while the tree settles.
+                engine.request_render_at(engine.get_time() + 16 * time.Millisecond)
             }
         }
 
@@ -105,6 +127,10 @@ draw_ui :: proc() {
 // Toast needs a Java looper on this thread. Preparing one is a one-time cost
 // and the native ALooper the engine polls is unaffected -- they're separate.
 show_toast :: proc(msg: string) {
+    // The bindings all fail closed without a VM, so this would be harmless off
+    // Android -- just a pointless chain of JNI calls on every tap. Compile it
+    // out instead of relying on that.
+    when ODIN_PLATFORM_SUBTARGET == .Android {
         if l, ok := looper_my_looper(); ok && l == nil {
             looper_prepare()
         }
@@ -114,11 +140,20 @@ show_toast :: proc(msg: string) {
             return
         }
         toast_show(toast)
+    }
 }
 
 // Walks Activity -> Window -> decor View -> WindowInsets. Returns false while
 // the decor view is still unattached, which is normal for the first few frames.
 refresh_insets :: proc() -> bool {
+    // Returning true, not false: zero insets are the correct answer where there
+    // are no system bars, and the caller retries every frame until this succeeds
+    // -- requesting a render each time. Failing here would pin the desktop loop
+    // awake forever, which is exactly what render_on_change is meant to avoid.
+    when ODIN_PLATFORM_SUBTARGET != .Android {
+        return true
+    }
+
     // This runs every frame until it succeeds, and each call below returns a
     // local ref. Without a frame to scope them the ref table fills and aborts.
     env := jni.env()