Release 1.4

versionCode 5. public_html/ changed, so the service worker cache goes to
bgclock-v20.

Confirmed on device: a turn left running well past the delay, force-quit and
reopened, now comes back with the time it had actually spent.

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

.gitignore

@@ -6,3 +6,6 @@ android/local.properties
 # editor leftovers
 *.swp
 *~
+
+# python leftovers
+__pycache__/

RELEASING.md

@@ -61,6 +61,14 @@ make some moves, force-quit from the task switcher and reopen — everything
 should come back, paused. That is the check that proves the app's storage is
 working, and it fails silently rather than loudly.
 
+**Let the turn run at least 40 seconds before you quit.** This is the whole
+check, and doing it any faster proves nothing: for the first 12 seconds of a
+turn the delay absorbs everything and the reserve clock is *supposed* to sit
+still, so a clock that has lost its place looks exactly like one that hasn't.
+Quit well past the delay and the time must come back lower than it was at the
+turn switch — and the delay counter must not be sitting at a full 12. That is
+how a broken save hid from this check from 1.0 to 1.3.
+
 ## 5. Commit, tag, push
 
 ```bash
@@ -110,10 +118,29 @@ Separate from all of the above, and manual: upload the six files in
 
 ## If the screenshots need redoing
 
-`fastlane/metadata/android/en-US/images/phoneScreenshots/` holds five shots at
-1170×2532. They are generated from `public_html/index.html` itself with headless
-Chromium, not taken on a phone, so they can be regenerated exactly. Two traps if
-you do it by hand:
+```bash
+python3 tools/screenshots.py
+```
+
+That rewrites all five in
+`fastlane/metadata/android/en-US/images/phoneScreenshots/` at 1170×2532. They
+are rendered from `public_html/index.html` with headless Chromium rather than
+taken on a phone, so anything that changes the look of the app is a reason to
+re-run it.
+
+**Near-exact, not exact.** Two runs of identical code differ by a few bytes in
+shots 1 and 4: a 4×13 pixel sliver at the delay bar's fill edge, where the
+`scaleX` boundary lands mid-pixel and the rasteriser rounds it differently from
+one run to the next. Max channel difference 11, invisible. So a `git diff` after
+re-running is not evidence that anything changed — compare the pixels before
+believing it, and `git checkout` the files back if that sliver is all it is.
+
+The script sets each scene by seeding `localStorage` before the app boots, so
+the app renders its own saved state rather than having the DOM poked from
+outside. Editing a scene means editing the `SCENES` table at the top.
+
+Three traps, all of which fail quietly, and all of which the script already
+handles — they are written down here because they cost an afternoon each:
 
 - The capture lands whenever the virtual-time budget runs out, not when your
   script finishes — so freeze the render loop (`requestAnimationFrame = () => 0`)
@@ -121,6 +148,10 @@ you do it by hand:
 - CSS transitions don't advance under `--virtual-time-budget`, so a panel caught
   mid-transition photographs in its *old* colour. Disable transitions in the
   screenshot build.
+- Snap-confined Chromium cannot write into hidden directories — not `~/.cache`,
+  not any dot-directory, not the private `/tmp` it is handed. It fails with a
+  bare "Permission denied" and no hint as to why. Both the HTML it reads and the
+  PNG it writes have to sit in a plainly-named directory.
 
 ## Reference
 

android/app/build.gradle.kts

@@ -12,8 +12,8 @@ android {
         applicationId = "nl.hansdezwart.bgclock"
         minSdk = 26
         targetSdk = 36
-        versionCode = 2
-        versionName = "1.1"
+        versionCode = 5
+        versionName = "1.4"
     }
 
     compileOptions {

android/app/src/main/java/nl/hansdezwart/bgclock/MainActivity.java

@@ -121,8 +121,15 @@ public class MainActivity extends Activity {
         }
     }
 
-    // forwarded so the page's visibilitychange fires — that is what calls
-    // saveGame() and nudges the audio context back after backgrounding
+    // Forwarded so the WebView stops its timers, animations and audio while the
+    // app is in the background, and picks them up again on return.
+    //
+    // This does NOT drive the page's visibilitychange — an earlier comment here
+    // claimed it did, and the saved match quietly depended on that being true.
+    // Document visibility follows the WebView's view and window visibility, not
+    // this call, so the page can be killed without ever hearing about it. The
+    // page no longer relies on being told: it writes the match to localStorage
+    // once a second while a clock is running.
     @Override
     protected void onPause() {
         super.onPause();

fastlane/metadata/android/en-US/changelogs/3.txt

@@ -0,0 +1 @@
+Improved the contrast to be compliant with accessibility guidelines.

fastlane/metadata/android/en-US/changelogs/4.txt

@@ -0,0 +1,4 @@
+Resetting the clock is now a press and hold rather than two taps. The ring
+around the button fills while you hold it, and springs back if you let go
+early, so the button explains itself. A short sound confirms the reset when
+sound is on.

fastlane/metadata/android/en-US/changelogs/5.txt

@@ -0,0 +1,4 @@
+Fixes a match not coming back correctly after the app is force-quit. The clock
+returned to where it stood at the last turn change, losing the time spent on
+the turn that was interrupted. The match is now written down once a second
+while a clock is running, so at most a second can be lost.

public_html/index.html

@@ -14,16 +14,18 @@
 <link rel="apple-touch-icon" href="icon-192.png">
 <style>
   :root{
-    --accent:#7E9A79;
+    --accent:#7C9777;
     --accent-ink:#FFFFFF;
-    --accent-mute:rgba(0,0,0,.34);
+    --accent-mute:rgba(0,0,0,.66);   /* replaced per theme by applyTheme() */
+    --track:rgba(0,0,0,.34);         /* the delay bar, not text — see .track */
     --idle:#9A9A9A;
     --idle-ink:#2E2E2E;
-    --idle-mute:rgba(0,0,0,.36);
+    --idle-mute:rgba(0,0,0,.63);
     --bar:#332E2B;
     --icon:#8E8880;
     --icon-hot:#DFD9D1;
     --flag:#A6564E;
+    --flag-mute:rgba(255,255,255,.82);
     --sheet:#232120;
     --sheet-line:rgba(255,255,255,.09);
     --sheet-ink:#EDE9E4;
@@ -90,7 +92,11 @@
                                lines up while both measure exactly the same */
     gap:3px 8px;
     order:-1;
-    font-size:clamp(13px,3.7vw,20px);
+    /* the 19px floor is load-bearing, not taste: WCAG's large-text threshold is
+       14pt bold = 18.66px, and below it these labels need 4.5:1 rather than 3:1
+       — which no ink still readable as "muted" can reach on a mid-tone accent.
+       Lower this and the contrast stops being conformant, silently. */
+    font-size:clamp(19px,3.9vw,24px);
     font-weight:700;
     letter-spacing:.01em;
     color:var(--idle-mute);
@@ -103,9 +109,8 @@
      The guard drops the space when there's no score line above to separate from. */
   .line:not([hidden]) + .line--moves > *{margin-top:7px}
 
-  .panel[data-state="active"] .head,
-  .panel[data-state="flagged"] .head{color:var(--accent-mute)}
-  .panel[data-state="flagged"] .head{color:rgba(255,255,255,.55)}
+  .panel[data-state="active"]  .head{color:var(--accent-mute)}
+  .panel[data-state="flagged"] .head{color:var(--flag-mute)}
 
   .stack{
     flex:1;
@@ -151,7 +156,9 @@
     width:clamp(72px,20vw,140px);
     height:clamp(4px,1.1vw,7px);
     border-radius:99px;
-    background:var(--accent-mute);
+    /* its own token: --accent-mute is text ink now and far darker than this bar
+       wants, and the two only ever agreed by accident */
+    background:var(--track);
     opacity:.45;
     overflow:hidden;
   }
@@ -162,9 +169,17 @@
     transform-origin:left center;
   }
   .dnum{
-    font-size:clamp(15px,4.4vw,26px);
+    font-size:clamp(19px,4.4vw,26px);   /* 19px for the same reason as .head */
     font-weight:700;
-    min-width:1.6em;
+    /* Exactly two digits — the delay caps at 60s, so it never needs a third —
+       and right-aligned, which does two things at once: the bar and the number
+       together sit dead centre of the panel whatever the count, and the digit
+       that changes every second stays put instead of sliding left when the
+       count drops out of double figures. The gap to the bar opens up by one
+       digit instead. 2ch is the digit advance, so with tabular figures this is
+       the exact width and not an approximation of it. */
+    min-width:2ch;
+    text-align:right;
   }
 
   /* ---------- control bar ---------- */
@@ -194,26 +209,73 @@
   .btn--vacant{visibility:hidden}
   .btn:focus-visible{outline:2px solid var(--icon-hot);outline-offset:-6px;border-radius:8px}
 
-  /* reset armed for a second tap — the ring unwinds over the same 2s as the timeout */
+  /* Reset confirms two different ways, because the two input methods can't share
+     one. A finger holds the button and the ring fills under it; lifting early
+     springs the ring back, which is the only way to say "keep going" without
+     words — and the bar carries no words by design, since it is shared by two
+     players sitting opposite each other. Keyboard and assistive activation
+     can't hold, so they keep the older two-presses-within-2s, where the ring
+     drains to show the window closing. */
   #reset{position:relative}
-  #reset.armed{color:var(--icon-hot)}
+  #reset.armed,#reset.holding{color:var(--icon-hot)}
   #reset .ring{
     position:absolute;
     top:50%;left:50%;
-    width:clamp(36px,9.6vw,48px);height:clamp(36px,9.6vw,48px);
+    /* As big as the bar will take, so a thumb can't cover it: the button is the
+       bar's full height, leaving 4px of dark above and below. width:auto is
+       load-bearing — .btn svg sets a width for the icons, and inheriting it here
+       would give the ring that width against this height and draw an ellipse.
+       It ends up a few px wider than the button's own slot, which is only ever
+       empty space; the neighbouring glyph stays a good 12px clear. */
+    height:calc(100% - 8px);
+    width:auto;
+    aspect-ratio:1;
     transform:translate(-50%,-50%) rotate(-90deg);   /* start the arc at 12 o'clock */
-    display:none;
+    opacity:0;
+    /* Hiding waits out the spring-back, then fades. display:none here would cut
+       the retreat off at the instant the finger lifts, and that retreat is the
+       whole reason the gesture teaches itself. The 120ms is a real duration on
+       purpose: a 0s transition with a delay is allowed to be treated as no
+       transition at all, which would take the delay with it. */
+    transition:opacity 120ms linear 180ms;
     pointer-events:none;
   }
-  #reset.armed .ring{display:block}
+  /* appearing is instant — no duration, no delay */
+  #reset.armed .ring,#reset.holding .ring{opacity:1;transition:none}
+
   #reset .ring circle{
     fill:none;
-    stroke:currentColor;                             /* the same flash the buttons use */
-    stroke-width:2.6;
+    /* user units, so it scales with the ring: 2 of 40 lands between 2.5 and
+       3.4 real px across phone sizes, a shade heavier than the icon strokes
+       because this one has to read past a thumb */
+    stroke-width:2;
+  }
+  /* the whole circle, dim, there from the moment the finger lands: it shows how
+     far there is to go, which an arc growing out of nothing cannot */
+  #reset .ring-track{stroke:var(--icon)}
+  /* and the bright arc that covers it, from 12 o'clock clockwise */
+  #reset .ring-fill{
+    stroke:var(--icon-hot);
     stroke-linecap:round;
     stroke-dasharray:113.1;                          /* 2πr, r=18 */
-    animation:unwind 2s linear forwards;
+    stroke-dashoffset:113.1;                         /* empty */
+    /* The spring back. 180ms is the ceiling, not the figure: reversing a
+       transition part-way shortens it in proportion, so letting go at 41% takes
+       74ms, and the arc never spends longer retreating than it spent filling.
+       The dim track outlasts it either way — the ring holds for 180ms and then
+       fades, so even a stray tap shows a whole circle appear and go, which is
+       what says there was something to finish. */
+    transition:stroke-dashoffset 180ms ease-out;
   }
+  #reset.holding .ring-fill{
+    stroke-dashoffset:0;
+    transition-duration:975ms;                       /* keep in step with HOLD_MS */
+    transition-timing-function:linear;
+  }
+  /* the keyboard path: a full ring draining over the 2s window. An animation
+     beats a transition on the same property, which is fine here only because
+     .armed and .holding are never set at once. */
+  #reset.armed .ring-fill{animation:unwind 2s linear forwards}
   @keyframes unwind{
     from{stroke-dashoffset:0}
     to{stroke-dashoffset:113.1}
@@ -347,6 +409,13 @@
 
   @media (prefers-reduced-motion:reduce){
     *{transition-duration:.01ms!important}
+    /* ...except the reset ring, which reports how far through the hold you are.
+       Collapsing it would fill the ring the instant you touched the button and
+       tell you it was done 650ms before it was. This one is state, not
+       decoration. Higher specificity than the * above, so it wins despite both
+       being !important. */
+    #reset .ring-fill{transition-duration:180ms!important}
+    #reset.holding .ring-fill{transition-duration:975ms!important}
   }
 </style>
 </head>
@@ -372,8 +441,12 @@
   </section>
 
   <div class="bar">
-    <button class="btn" id="reset" aria-label="Reset clock">
-      <svg class="ring" viewBox="0 0 40 40" aria-hidden="true"><circle cx="20" cy="20" r="18"/></svg>
+    <button class="btn" id="reset" aria-label="Reset clock, hold to confirm">
+      <!-- ring-track/ring-fill, not track/fill: those two are the delay bar's -->
+      <svg class="ring" viewBox="0 0 40 40" aria-hidden="true">
+        <circle class="ring-track" cx="20" cy="20" r="18"/>
+        <circle class="ring-fill" cx="20" cy="20" r="18"/>
+      </svg>
       <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round">
         <path d="M20 6v6h-6"/><path d="M20 12a8 8 0 1 0-2.3 5.3"/>
       </svg>
@@ -584,12 +657,17 @@
   // ask the browser not to evict us when the device is low on space
   try{ if(navigator.storage && navigator.storage.persist) navigator.storage.persist(); }catch(e){}
 
+  /* mute is the alpha the head labels need over that accent to clear 4:1. It
+     can't be one shared value: the accents differ in lightness by more than the
+     muting does, and a single alpha either fails on plum or blackens brass.
+     sage and brass are a touch darker than they look elsewhere — white on them
+     has to clear 3:1 for the clock digits, and the old brass managed 2.46. */
   var THEMES = [
-    { id:"sage",   accent:"#7E9A79", ink:"#FFFFFF" },
-    { id:"brass",  accent:"#C3A153", ink:"#FFFFFF" },
-    { id:"slate",  accent:"#6E8598", ink:"#FFFFFF" },
-    { id:"teal",   accent:"#5E8D87", ink:"#FFFFFF" },
-    { id:"plum",   accent:"#8A7396", ink:"#FFFFFF" }
+    { id:"sage",   accent:"#7C9777", ink:"#FFFFFF", mute:.66 },
+    { id:"brass",  accent:"#AA8C47", ink:"#FFFFFF", mute:.66 },
+    { id:"slate",  accent:"#6E8598", ink:"#FFFFFF", mute:.72 },
+    { id:"teal",   accent:"#5E8D87", ink:"#FFFFFF", mute:.71 },
+    { id:"plum",   accent:"#8A7396", ink:"#FFFFFF", mute:.77 }
   ];
 
   var cfg = {
@@ -697,6 +775,15 @@
     tone(659.25, 0.16, 0.20, 0.24, "triangle", false);
     tone(440, 0.32, 0.55, 0.26, "triangle", true);
   }
+  // the flag's mirror: the same triad rising, and quicker — a fresh start.
+  // Deliberately not another rising fifth; sndSwap owns that shape and plays on
+  // every single handover, so nothing else may sound like it.
+  function sndReset(){
+    if(!cfg.sound) return;
+    tone(523.25, 0.00, 0.16, 0.20, "triangle", false);
+    tone(659.25, 0.07, 0.16, 0.20, "triangle", false);
+    tone(783.99, 0.14, 0.40, 0.22, "triangle", true);
+  }
 
   /* ============ helpers ============ */
   function elapsed(){
@@ -905,16 +992,19 @@
 
   // clocks back to full for the match length; the score is not this function's business
   function resetClocks(){
-    disarm();
+    cancelHold();
     st.reserve = [matchTime(), matchTime()];
     newGame();
     saveGame();
   }
 
-  // the Reset button — a whole new match
+  // the Reset button — a whole new match. The sound lives here rather than in
+  // resetClocks(), which the settings sheet also calls: changing the match
+  // length shouldn't sound like a reset.
   function resetAll(){
     st.score = [0, 0];
     resetClocks();
+    sndReset();
   }
 
   function clampScore(){
@@ -961,26 +1051,88 @@
     if(st.phase === PAUSE && st.active < 0) st.phase = IDLE;   // can't be mid-turn with no turn
   }
 
-  // the only hooks iOS reliably fires before killing an app
+  // The hooks iOS fires before killing an app. Android's WebView is the problem
+  // child: neither of these is guaranteed there, whatever MainActivity's
+  // onPause() forwarding suggests, and when they don't fire the newest save is
+  // the one tap() wrote — which carries held ≈ 0, because tap() sets turnAt and
+  // saves in the same breath. Reopening then hands back the entire turn, and
+  // the bug hides in plain sight: quit inside the delay window and a clock that
+  // correctly hasn't moved looks exactly like a clock that lost its place.
   document.addEventListener("visibilitychange", function(){
     if(document.visibilityState === "hidden") saveGame();
   });
   window.addEventListener("pagehide", saveGame);
+  // a different signal from visibilitychange, and it does fire when Android's
+  // task switcher opens. saveGame() is idempotent, so an extra call is free
+  window.addEventListener("blur", saveGame);
 
-  /* reset needs two taps within 2s — one stray tap can't wipe a live game */
-  var armT = null;
-  function disarm(){
+  // and the one that doesn't depend on being told anything: a second is the
+  // most a force-quit can now cost, which is well inside the delay nobody is
+  // charged for. A timer rather than a hook in loop(), because rAF stops when
+  // the page is hidden — exactly when this matters most.
+  setInterval(function(){ if(st.phase === RUN) saveGame(); }, 1000);
+
+  /* Reset is a hold, not a tap. It used to be two taps inside 2s, and a test
+     user found the button, tapped it once and gave up — which is exactly what
+     the old feedback invited, since a draining ring reads as "wait", not "press
+     me again". There is no wordless way to say "do that again"; there is a very
+     well-worn one for "keep holding", and letting go early springs the ring
+     back, which teaches the gesture without a word in the bar.
+
+     A stray 650ms hold is also far less likely than two stray taps in 2s, so
+     the live game it was guarding is better guarded than before. */
+  var HOLD_MS = 975;                // keep in step with the .holding transition
+  var SLOP = 12;                    // px a press may drift and still count; also the panels'
+  var holdT = null, armT = null, holdDown = null;
+
+  function cancelHold(){
+    if(holdT){ clearTimeout(holdT); holdT = null; }
     if(armT){ clearTimeout(armT); armT = null; }
+    holdDown = null;
+    btnReset.classList.remove("holding");
     btnReset.classList.remove("armed");
-    btnReset.setAttribute("aria-label", "Reset clock");
+    btnReset.setAttribute("aria-label", "Reset clock, hold to confirm");
   }
+
+  btnReset.addEventListener("pointerdown", function(ev){
+    ev.preventDefault();
+    audio();                        // unlocking on the press is a valid gesture, and earlier is better
+    cancelHold();
+    holdDown = { id: ev.pointerId, x: ev.clientX, y: ev.clientY };
+    // so the lift still lands here if the finger drifts off the button
+    try{ btnReset.setPointerCapture(ev.pointerId); }catch(e){}
+    btnReset.classList.add("holding");
+    holdT = setTimeout(function(){
+      cancelHold();                 // before, not after: resetAll() hides this button
+      resetAll();
+    }, HOLD_MS);
+  });
+
+  btnReset.addEventListener("pointermove", function(ev){
+    var d = holdDown;
+    if(!d || ev.pointerId !== d.id) return;
+    // a finger that has travelled this far was on its way somewhere else
+    if(Math.abs(ev.clientX - d.x) > SLOP || Math.abs(ev.clientY - d.y) > SLOP) cancelHold();
+  });
+
+  function endHold(ev){
+    if(holdDown && ev.pointerId !== holdDown.id) return;
+    cancelHold();                   // the spring back is the whole lesson
+  }
+  btnReset.addEventListener("pointerup", endHold);
+  btnReset.addEventListener("pointercancel", endHold);
+
+  /* The keyboard and switch path, which can't hold: two presses inside 2s, as
+     the whole button used to work. detail is 0 only for activation that didn't
+     come from a pointer — touch and mouse were dealt with above, and the click
+     trailing their press must not arm anything. */
   function armReset(){
-    if(armT){ disarm(); resetAll(); return; }             // second tap inside the window
+    if(armT){ cancelHold(); resetAll(); return; }       // second press inside the window
     btnReset.classList.remove("armed");
     void btnReset.offsetWidth;                          // restart the drain animation
     btnReset.classList.add("armed");
-    btnReset.setAttribute("aria-label", "Tap again to reset");
-    armT = setTimeout(disarm, 2000);
+    btnReset.setAttribute("aria-label", "Press again to reset");
+    armT = setTimeout(cancelHold, 2000);
   }
 
   /* ============ theme ============ */
@@ -988,6 +1140,7 @@
     var t = THEMES.filter(function(x){ return x.id === cfg.theme; })[0] || THEMES[0];
     document.documentElement.style.setProperty("--accent", t.accent);
     document.documentElement.style.setProperty("--accent-ink", t.ink);
+    document.documentElement.style.setProperty("--accent-mute", "rgba(0,0,0," + t.mute + ")");
     [].forEach.call(document.querySelectorAll(".sw"), function(s){
       s.setAttribute("aria-checked", s.dataset.id === cfg.theme ? "true" : "false");
     });
@@ -1114,7 +1267,7 @@
     // can reach the timer controls once the match is under way
     if(!matchPristine()) return;
     pauseNow();                 // nobody's clock runs while a sheet is open
-    disarm();
+    cancelHold();
     paintSettings();
     openSheet(sheet);
   });
@@ -1154,7 +1307,7 @@
     // set themselves survives
     resumeOnClose = (st.phase === RUN);
     pauseNow();
-    disarm();
+    cancelHold();
     // seed from what the panels are showing, not from the banked reserve — for
     // the interrupted turn those differ by the delay already spent
     edit   = { score: st.score.slice(), time: [shownTime(0), shownTime(1)] };
@@ -1209,7 +1362,10 @@
     closeScore();
   });
 
-  btnReset.addEventListener("click", armReset);
+  btnReset.addEventListener("click", function(ev){
+    if(ev.detail !== 0) return;   // the tail of a press we already handled
+    armReset();
+  });
   btnPlay.addEventListener("click", togglePause);
 
   /* ============ panel taps ============ */
@@ -1217,8 +1373,8 @@
   // would steal the turn from Android's swipe-up-from-the-bottom gesture: the
   // bottom panel owns that edge, and starting a swipe home is a finger landing.
   // A swipe travels past SLOP, or the system claims it and we get pointercancel
-  // instead of a lift. Either way there's no turn change.
-  var SLOP = 12;                  // px of travel a press may drift and still count
+  // instead of a lift. Either way there's no turn change. (SLOP is declared
+  // with the reset hold, which measures drift the same way.)
   var wantFullscreen = true;
   panels.forEach(function(p){
     var down = null;

public_html/sw.js

@@ -1,5 +1,5 @@
 /* Cache-first: once installed the clock never touches the network again. */
-const CACHE = "bgclock-v17";
+const CACHE = "bgclock-v20";
 const FILES = [
   "./",
   "./index.html",

tools/screenshots.py

@@ -0,0 +1,182 @@
+#!/usr/bin/env python3
+"""Regenerate the five F-Droid phone screenshots from public_html/index.html.
+
+    python3 tools/screenshots.py
+
+They are rendered rather than photographed, so they can be redone whenever the
+UI changes. Output is 1170x2532 — 390x844 CSS pixels at device scale 3, an
+iPhone-sized frame that F-Droid is happy with.
+
+Near-exact, not exact: two runs of identical code differ by a 4x13 pixel sliver
+at the delay bar's fill edge, where the scaleX boundary lands mid-pixel and the
+rasteriser rounds it differently. Invisible, but it means a non-empty git diff
+after re-running proves nothing on its own.
+
+Three things bite, and all three fail quietly rather than loudly:
+
+1. The capture lands when Chromium's virtual-time budget runs out, not when
+   this script finishes. So the page is frozen deliberately — the app's render
+   loop is `render(); requestAnimationFrame(loop)`, and replacing
+   requestAnimationFrame after the first synchronous render pins the frame.
+
+2. CSS transitions don't advance under --virtual-time-budget. A panel caught
+   mid-transition photographs in its *old* colour, so transitions are disabled
+   outright in the copy being shot.
+
+3. Snap-confined Chromium cannot write into hidden directories — not ~/.cache,
+   not a dot-directory anywhere, and not the private /tmp it gets given. It
+   fails with a bare "Permission denied" and no hint. Both the HTML it reads
+   and the PNG it writes therefore live in a plain directory in the repo.
+
+The scenes are set up by seeding localStorage before the app boots, so the app
+renders its own state from its own save format rather than having the DOM
+poked from outside. Only the final clicks (open a sheet, start the clock) are
+driven through the real buttons.
+"""
+
+import json
+import shutil
+import subprocess
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent.parent
+SOURCE = ROOT / "public_html" / "index.html"
+OUT = ROOT / "fastlane" / "metadata" / "android" / "en-US" / "images" / "phoneScreenshots"
+WORK = ROOT / "build-screenshots"          # plain name: see note 3 above
+
+WIDTH, HEIGHT, SCALE = 390, 844, 3         # -> 1170x2532
+
+CHROMIUM = ("chromium-browser", "chromium", "google-chrome", "google-chrome-stable")
+
+# st.phase in the app; a running clock is saved as PAUSE, because time that
+# passes while the app is dead can't be charged to anyone
+IDLE, RUN, PAUSE, FLAG = 0, 1, 2, 3
+
+MIN = 60_000
+
+# A five-point match, mid-game, from the bottom player's side. Panel i shows
+# score[i] as "You" and score[1-i] as "Them", so one array serves both ends.
+MATCH = dict(
+    cfg=dict(points=5, base=3 * MIN, delay=12_000),
+    game=dict(score=[1, 2], reserve=[14 * MIN, 13 * MIN + 15_000],
+              moves=[7, 6], active=1, held=5_000, phase=PAUSE),
+)
+
+SCENES = [
+    # 1 — the clock itself, mid-match, delay running down on the bottom player
+    dict(name="1", click="playpause", **MATCH),
+
+    # 2 — the score sheet, where a finished game is written down
+    dict(name="2", click="score", **MATCH),
+
+    # 3 — the settings sheet. It only opens on a pristine match (the app refuses
+    #     once a game is under way), so this one starts from a fresh state and
+    #     seeds no saved game at all.
+    dict(name="3", click="settings",
+         cfg=dict(points=5, base=3 * MIN, delay=12_000), game=None),
+
+    # 4 — a single game: no match score, so the score lines aren't there
+    dict(name="4", click="playpause",
+         cfg=dict(points=1, base=3 * MIN, delay=12_000),
+         game=dict(score=[0, 0], reserve=[3 * MIN, 3 * MIN],
+                   moves=[4, 3], active=1, held=5_000, phase=PAUSE)),
+
+    # 5 — a flagged clock: the top player's time is gone
+    dict(name="5", click=None,
+         cfg=dict(points=5, base=3 * MIN, delay=12_000),
+         game=dict(score=[1, 2], reserve=[0, 42_000],
+                   moves=[7, 6], active=0, held=0, phase=FLAG)),
+]
+
+# Runs in <head>, before the app's own script, so the app boots into this state
+# and renders it itself.
+SEED = """
+<script>
+(function(){
+  try{ localStorage.clear(); }catch(e){}
+  var seed = %s;
+  for(var k in seed) localStorage.setItem(k, seed[k]);
+})();
+</script>
+"""
+
+# Runs after the app's script, which ends with render(); loop().
+FREEZE = """
+<style>
+  /* transitions don't advance under --virtual-time-budget, so a panel caught
+     mid-transition would photograph in its old colour */
+  *, *::before, *::after{transition:none !important; animation:none !important}
+</style>
+<script>
+(function(){
+  // the first render has already happened synchronously; pin it
+  window.requestAnimationFrame = function(){ return 0; };
+  var click = %s;
+  if(click){
+    var b = document.getElementById(click);
+    if(!b) throw new Error("no such button: " + click);
+    b.click();
+  }
+  document.title = "shot-ready";
+})();
+</script>
+"""
+
+
+def find_chromium():
+    for name in CHROMIUM:
+        path = shutil.which(name)
+        if path:
+            return path
+    sys.exit("no chromium found; tried: " + ", ".join(CHROMIUM))
+
+
+def build_page(scene, source):
+    """A copy of the app with the scene seeded and the render loop pinned."""
+    seed = {
+        "bg.points": str(scene["cfg"]["points"]),
+        "bg.base": str(scene["cfg"]["base"]),
+        "bg.delay": str(scene["cfg"]["delay"]),
+        "bg.theme": scene["cfg"].get("theme", "sage"),
+        "bg.sound": "1",
+    }
+    if scene["game"] is not None:
+        seed["bg.game"] = json.dumps(dict(v=1, **scene["game"]))
+
+    html = source.replace("</head>", SEED % json.dumps(seed) + "</head>", 1)
+    html = html.replace("</body>", FREEZE % json.dumps(scene["click"]) + "</body>", 1)
+    return html
+
+
+def main():
+    chromium = find_chromium()
+    source = SOURCE.read_text()
+    WORK.mkdir(exist_ok=True)
+    try:
+        for scene in SCENES:
+            page = WORK / ("scene-%s.html" % scene["name"])
+            page.write_text(build_page(scene, source))
+            target = OUT / ("%s.png" % scene["name"])
+            shot = WORK / target.name
+
+            subprocess.run([
+                chromium,
+                "--headless", "--no-sandbox", "--disable-gpu", "--hide-scrollbars",
+                "--window-size=%d,%d" % (WIDTH, HEIGHT),
+                "--force-device-scale-factor=%d" % SCALE,
+                "--virtual-time-budget=4000",
+                "--screenshot=%s" % shot,
+                page.as_uri(),
+            ], check=True, capture_output=True)
+
+            if not shot.exists():
+                sys.exit("chromium wrote nothing for scene %s" % scene["name"])
+            shutil.move(str(shot), str(target))
+            print("%s  %d bytes" % (target.relative_to(ROOT), target.stat().st_size))
+    finally:
+        shutil.rmtree(WORK, ignore_errors=True)
+
+
+if __name__ == "__main__":
+    main()