Stop the kiosk coming up blank after every reboot

A reboot kills Chromium rather than closing it, so it writes "exit_type":"Crashed" into
its profile. On the next boot it came up with a restore prompt and a tab stuck for ever
on a blank "Loading..." page - navigation never committed, navigator.serviceWorker was
still undefined, nothing on screen. On a device with a monitor and no keyboard that is
permanent, and the backend was healthy throughout: / answered in 5 ms and /api/library
returned 812 kB the whole time.

--disable-session-crashed-bubble hides the prompt but not the state behind it, and here
the prompt showed anyway. .xinitrc now rewrites exit_type and exited_cleanly before
launching, which is the usual remedy for a kiosk that gets powered off rather than shut
down, and adds --hide-crash-restore-bubble (the current name for the flag).

Verified by rebooting and leaving it alone: the page comes up on its own with its title
and its content.

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

roles/pi_kiosk/templates/xinitrc.j2

@@ -29,6 +29,18 @@ BROWSER_ARGS="--use-gl=egl --enable-gpu-rasterization --ignore-gpu-blocklist --e
 BROWSER_ARGS="$BROWSER_ARGS --noerrdialogs --disable-infobars --no-first-run"
 BROWSER_ARGS="$BROWSER_ARGS --no-default-browser-check --disable-session-crashed-bubble"
 BROWSER_ARGS="$BROWSER_ARGS --password-store=basic --check-for-update-interval=31536000"
+BROWSER_ARGS="$BROWSER_ARGS --hide-crash-restore-bubble"
+
+# Chromium records in its profile how it last exited. A reboot kills it rather than
+# closing it, so that record says "Crashed", and the next start comes up with a restore
+# prompt and a tab stuck on a blank "Loading..." page - which, on a device with no
+# keyboard in front of it, it never leaves. The flags above hide the prompt but do not
+# clear the state behind it; rewriting the two keys before launch does, and is the
+# standard remedy for a kiosk that is powered off rather than shut down.
+PREFS="$HOME/.config/chromium/Default/Preferences"
+if [ -f "$PREFS" ]; then
+  sed -i 's/"exit_type":"[^"]*"/"exit_type":"Normal"/; s/"exited_cleanly":false/"exited_cleanly":true/' "$PREFS"
+fi
 {% if pi_kiosk_mode == 'debug' %}
 # Debug mode only. Binds to 127.0.0.1, so it is reachable over an ssh tunnel and from
 # nowhere else. Lets the page be driven and measured from a shell - `curl

roles/pi_musicmouse/handlers/main.yml

@@ -26,6 +26,15 @@
     src: "{{ pi_musicmouse_build_cache }}/{{ musicmouse_version | regex_replace('/', '_') }}/web/dist/"
     dest: /opt/musicmouse/web/dist/
     delete: true
+  notify: Record which commit the frontend was built from
+
+- name: Record which commit the frontend was built from
+  # After the sync, and outside dist/, which the sync empties of anything it did not
+  # build. This is what lets the next run tell a current dist from a stale one.
+  ansible.builtin.copy:
+    content: "{{ pi_musicmouse_web_checkout.after }}\n"
+    dest: /opt/musicmouse/web/.built-commit
+    mode: "0644"
   notify: Restart musicmouse
 
 - name: Restart musicmouse

roles/pi_musicmouse/tasks/main.yml

@@ -194,21 +194,34 @@
   become: false
   vars:
     ansible_become: false  # see the first delegated task in tasks/main.yml
+  register: pi_musicmouse_web_checkout
   notify: Build frontend
 
-- name: Check whether the device has a built frontend
+- name: Read which commit the frontend on the device was built from
-  ansible.builtin.stat:
+  # Same guard as the venv's .installed-commit above, and for the same reason. Hanging
-    path: /opt/musicmouse/web/dist/index.html
+  # the build chain off the control machine's checkout changing has a hole: a run that
-  register: pi_musicmouse_dist
+  # fails after the checkout leaves the device's dist as it was, and every later run
+  # finds the checkout current and rebuilds nothing, while ansible reports no changes.
+  # Stale is the dangerous case - a missing dist is at least obvious - and it is exactly
+  # what happens after a dist is replaced by hand, which is easy to do while chasing a
+  # performance problem on the device.
+  #
+  # The stamp lives beside dist/ rather than inside it: `Sync frontend` rsyncs with
+  # --delete, so anything in dist/ that the build did not produce is removed.
+  ansible.builtin.command:
+    cmd: cat /opt/musicmouse/web/.built-commit
+  register: pi_musicmouse_web_stamp
+  changed_when: false
+  failed_when: false
 
-- name: Queue a frontend build when the device has none
+- name: Queue a frontend build when the device's is missing or from another commit
-  # The build chain hangs off the control machine's checkout changing, which has the
-  # same hole the backend install had: a run that fails after that checkout leaves the
-  # Pi with no dist, and every later run finds the checkout current and rebuilds
-  # nothing. Cheap to check, and it only fires when the device really has no UI.
   ansible.builtin.debug:
-    msg: "No frontend at /opt/musicmouse/web/dist - queueing a build"
+    msg: >-
-  when: not pi_musicmouse_dist.stat.exists
+      Frontend on the device is at
+      {{ pi_musicmouse_web_stamp.stdout | default('(none)', true) | trim }},
+      wanted {{ pi_musicmouse_web_checkout.after }} - queueing a build
+  when: pi_musicmouse_web_stamp.stdout | default('', true) | trim
+        != pi_musicmouse_web_checkout.after
   changed_when: true
   notify: Build frontend