diff --git a/PROJECT.md b/PROJECT.md
index 558a5d6..3f73d0c 100644
--- a/PROJECT.md
+++ b/PROJECT.md
@@ -2582,15 +2582,40 @@ inside a `<script type="module">` is a hard 404 the moment the same file is serv
   purpose — iOS already routes these through the system's own vibration setting.
   **Untested on hardware.** The simulator has no Taptic Engine, so every weight and gap below is a
   guess on paper and needs its own phone round before any of it is assumed good.
-- **SWIPE LEFT GOES BACK** in the menus. Every screen but the splash wears exactly one chevron, and
-  that button already knows where back is for its own screen — so the gesture finds the visible
-  `.chev` and clicks it rather than keeping a second map of the wizard that would drift the first
-  time a stage is added. Threshold 60 px, horizontal-dominant so the stats and board screens still
-  scroll. The `touchend` listener is deliberately **not passive**: a swipe that starts and ends
-  inside one of the big poster cards also produces that card's click, and `preventDefault` is what
-  stops the screen answering forward in the same frame the chevron walked it back.
-  **iOS convention is the opposite** (the system pop gesture drags rightward from the left edge).
-  Built to the brief as asked; flipping it is the sign on one comparison in `main.js`.
+- **SWIPE RIGHT GOES BACK** in the menus — the direction iOS's own pop gesture goes. (Built leftward
+  first, to the letter of the brief; Oliver flipped it the same hour: *"that's the iOS way"*. The
+  lesson is cheap and worth keeping: on a platform gesture, the platform's convention beats the
+  literal wording of the request — say so up front rather than shipping it and flagging it after.)
+  Every screen but the splash wears exactly one chevron, and that button already knows where back is
+  for its own screen — so the gesture finds the visible `.chev` and clicks it rather than keeping a
+  second map of the wizard that would drift the first time a stage is added. Threshold 60 px,
+  horizontal-dominant so the stats and board screens still scroll. The `touchend` listener is
+  deliberately **not passive**: a swipe that starts and ends inside one of the big poster cards also
+  produces that card's click, and `preventDefault` is what stops the screen answering forward in the
+  same frame the chevron walked it back.
+- **SOUND, and it was broken twice over on the first shell build.** Oliver, minutes after it ran:
+  *"I don't hear any of the games sounds.. daniel the umpire is synthesized which is horrible and no
+  music is playing on start"*. Two independent causes, both now fixed, both worth knowing before the
+  next plugin:
+  1. **`fetch` responses from the app bundle are OPAQUE.** Capacitor serves `capacitor://localhost`
+     through a custom `WKURLSchemeHandler`, and `fetch` reports those as `status: 0, ok: false` —
+     with the full body present and decoding perfectly. `audio.js` guarded all three of its loads
+     with a plain `if (!r.ok) throw`, so inside the app EVERY recorded sound failed at the door: the
+     umpire fell through to the synthesized `say()`, no bed ever started, no sample ever played.
+     The synthesized sfx kept working, which made it read as a mixing bug rather than a loader one.
+     Fixed once, in the shared `bytes()` helper every audio fetch now routes through: **a status of
+     0 is the absence of an HTTP status, not a failure** — trust the status only when there is one,
+     so a real 404 still falls back. **Any future `!r.ok` guard on a bundle-relative fetch is this
+     same bug.** (`stats.js` is fine: its board calls go to the live host over real HTTPS.)
+  2. **The ring switch muted everything on device.** WKWebView's WebAudio runs in the SoloAmbient
+     session, which the hardware silent switch kills outright — a completely silent game on a phone
+     that lives on silent, with nothing on screen to explain it. `AppDelegate` now sets
+     `AVAudioSession .playback`, the games-and-media category. That is a shell-only fix; the web
+     build has no say in it, and it is a good part of why the shell is worth having for sound.
+
+  Confirmed in the shell after both fixes: the bed reaches `PLAYING`, the context is `running`
+  **with no gesture at all** (autoplay really is allowed in WKWebView, as LATER predicted), and the
+  `say()` fallback never fires. On the web, a genuinely missing file still falls back — checked.
 - **The app has no server behind it.** Inside the shell the page is served from the bundle
   (`capacitor://localhost`), so `fetch('api/scores')` and the netplay socket would have resolved
   against the app's own files. `config.NATIVE`/`config.API` and one branch in `net.js` point the
diff --git a/index.html b/index.html
index 174f8da..1906c10 100644
--- a/index.html
+++ b/index.html
@@ -661,6 +661,6 @@ if (location.hostname !== 'splatter-tennis.wuwi.one' && !window.Capacitor?.isNat
     </div>
   </div>
 </div>
-<script type="module" src="src/main.js?v=296"></script>
+<script type="module" src="src/main.js?v=297"></script>
 </body>
 </html>
diff --git a/ios/App/App.xcodeproj/project.pbxproj b/ios/App/App.xcodeproj/project.pbxproj
index b0cd4e4..75fa23c 100644
--- a/ios/App/App.xcodeproj/project.pbxproj
+++ b/ios/App/App.xcodeproj/project.pbxproj
@@ -349,6 +349,7 @@
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
+				DEVELOPMENT_TEAM = T789785GZK;
 				INFOPLIST_FILE = App/Info.plist;
 				IPHONEOS_DEPLOYMENT_TARGET = 14.0;
 				LD_RUNPATH_SEARCH_PATHS = "$(inherited) @executable_path/Frameworks";
@@ -369,6 +370,7 @@
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
+				DEVELOPMENT_TEAM = T789785GZK;
 				INFOPLIST_FILE = App/Info.plist;
 				IPHONEOS_DEPLOYMENT_TARGET = 14.0;
 				LD_RUNPATH_SEARCH_PATHS = "$(inherited) @executable_path/Frameworks";
diff --git a/ios/App/App/AppDelegate.swift b/ios/App/App/AppDelegate.swift
index c3cd83b..174f0b7 100644
--- a/ios/App/App/AppDelegate.swift
+++ b/ios/App/App/AppDelegate.swift
@@ -1,5 +1,6 @@
 import UIKit
 import Capacitor
+import AVFoundation
 
 @UIApplicationMain
 class AppDelegate: UIResponder, UIApplicationDelegate {
@@ -7,7 +8,14 @@ class AppDelegate: UIResponder, UIApplicationDelegate {
     var window: UIWindow?
 
     func application(_ application: UIApplication, didFinishLaunchingWithOptions launchOptions: [UIApplication.LaunchOptionsKey: Any]?) -> Bool {
-        // Override point for customization after application launch.
+        // PLAY THROUGH THE RING SWITCH. WKWebView's WebAudio runs in the SoloAmbient session by
+        // default, which the hardware silent switch mutes outright — so a phone that lives on silent
+        // (most of them) gets a completely silent game and nothing on screen explains why. `.playback`
+        // is the games-and-media category: audible on silent, and it is what makes the umpire and the
+        // menu theme reach a real device at all. The web build has no say in this; it is the whole
+        // reason the shell is worth having for sound. See PROJECT.md, "What the shell buys".
+        try? AVAudioSession.sharedInstance().setCategory(.playback)
+        try? AVAudioSession.sharedInstance().setActive(true)
         return true
     }
 
diff --git a/src/audio.js b/src/audio.js
index 17b6a52..7a37b9a 100644
--- a/src/audio.js
+++ b/src/audio.js
@@ -7,6 +7,20 @@ export const audio = () => {
   return AC;
 };
 
+// EVERY audio file is fetched through here, and the reason is one line of it. Inside the Capacitor
+// shell the bundle is served by a custom WKURLSchemeHandler, and `fetch` reports those responses as
+// OPAQUE: `status` 0 and `ok` false, even though the bytes are all present and decode perfectly.
+// The old guard was a plain `if (!r.ok) throw` — so in the app EVERY recorded sound failed at the
+// door: the umpire fell through to the synthesized speech fallback, no bed ever started, and not a
+// single sample played, while the synthesized sfx kept working and made it look like a mixing bug
+// (Oliver, 2026-08-20, on the first shell build: *"daniel the umpire is synthesized which is
+// horrible and no music is playing on start"*). A status of 0 is not a failure, it is the absence
+// of an HTTP status — so trust the status only when there IS one, and let a real 404 still throw.
+const bytes = url => fetch(url).then(r => {
+  if (r.status && !r.ok) throw new Error(url + ' ' + r.status);
+  return r.arrayBuffer();
+});
+
 // Umpire lines are pre-recorded files (assets/audio/*.m4a) played via WebAudio —
 // works on mobile where speechSynthesis is unreliable. say() remains the fallback.
 const VOICE = {};
@@ -48,8 +62,7 @@ export function voice(name, delay = 0, rate = 1) {
 
 function loadVoice(name) {
   if (VOICE[name]) return Promise.resolve(VOICE[name]);
-  return fetch('assets/audio/' + name + '.m4a')
-    .then(r => { if (!r.ok) throw 0; return r.arrayBuffer(); })
+  return bytes('assets/audio/' + name + '.m4a')
     .then(ab => audio().decodeAudioData(ab))
     .then(buf => (VOICE[name] = buf));
 }
@@ -119,8 +132,7 @@ export function sfx(name, vol = 1, opts = {}) {
   };
   if (SFX[name]) return play(SFX[name]);
   // a name with an extension is used as-is (weapon-click2.mp3); bare names keep the .wav default
-  fetch('assets/audio/' + (name.includes('.') ? name : name + '.wav'))
-    .then(r => { if (!r.ok) throw 0; return r.arrayBuffer(); })
+  bytes('assets/audio/' + (name.includes('.') ? name : name + '.wav'))
     .then(ab => ac.decodeAudioData(ab))
     .then(buf => { SFX[name] = buf; play(buf); })
     .catch(() => {}); // missing sample = silence, never a thrown error mid-rally
@@ -141,8 +153,7 @@ function loopBed(file, lowcut, vol, loopEnd = 0) {
   const ac = audio();
   if (beds.has(file)) return;
   beds.set(file, null);
-  fetch('assets/audio/' + file) // full filename: the beds are a mix of .wav and .m4a, see below
-    .then(r => { if (!r.ok) throw 0; return r.arrayBuffer(); })
+  bytes('assets/audio/' + file) // full filename: the beds are a mix of .wav and .m4a, see below
     .then(ab => ac.decodeAudioData(ab))
     .then(buf => {
       // RESUME AGAIN, HERE. The gesture that unlocked the context is long gone by now: the menu
diff --git a/src/main.js b/src/main.js
index ccca17e..cfa782c 100644
--- a/src/main.js
+++ b/src/main.js
@@ -537,7 +537,8 @@ document.getElementById('sBack').addEventListener('click', () => {
 });
 document.getElementById('bsBack').addEventListener('click', () => go('secret'));
 
-// SWIPE TO GO BACK. Every screen but the splash wears exactly one chevron in its header (.hdr in
+// SWIPE TO GO BACK — RIGHTWARD, the way iOS's own pop gesture goes (Oliver, 2026-08-20: *"that's
+// the iOS way"*). Every screen but the splash wears exactly one chevron in its header (.hdr in
 // index.html) and that button ALREADY knows where back is for its own screen — the wizard's shape
 // lives in the handlers above and nowhere else. So the gesture finds the one visible chevron and
 // clicks it, rather than keeping a second map that would drift the first time a stage is added.
@@ -553,8 +554,8 @@ intro.addEventListener('touchend', e => {
   if (!from || e.touches.length) return;
   const t = e.changedTouches[0];
   const dx = t.clientX - from.x, dy = t.clientY - from.y;
-  // leftward, and clearly horizontal: the stats and board screens scroll vertically under the finger
-  if (dx > -SWIPE_MIN || Math.abs(dy) > Math.abs(dx) * 0.7) return;
+  // rightward, and clearly horizontal: the stats and board screens scroll vertically under the finger
+  if (dx < SWIPE_MIN || Math.abs(dy) > Math.abs(dx) * 0.7) return;
   const back = [...intro.querySelectorAll('.chev')].find(b => b.offsetParent); // the one on screen
   if (!back) return;
   // NOT passive, and this is why: a swipe that starts and ends inside one of the big poster cards