smoke: bump the kumaui pin and re-arm the two wave-era probes

kumaui's idle-without-windows fix (e90884c) keeps the shell alive
holding its bus names and sleep inhibitor through output loss, with
surfaces recreated when displays return — noctalia's contract, which
is what the CI VM has been measuring the absence of. The pin resolves
against kumaui main now pushed (6bf413b: the fix, the vendor-doc note
under it, and the memory-seam commit on top; the diff is the review).

The notification-ownership probe and the sleep-inhibitor readback come
back — polled, because the unit still starts seconds after loginctl
sees the session — and the inhibitor grep names kuma-shell, not
noctalia, for the first time. The run that carries this is the
independent verification of the kumaui claim.

Author: Letdown2491

.github/actions/kuma-shell/action.yml

@@ -33,7 +33,7 @@ inputs:
     description: >-
       kumaui commit to build the shell from.
     required: false
-    default: 764c850f1fb1c13c0e396e279f555e1db3a3078d
+    default: 6bf413bf81b07fd71126292c978ca35f9ca3c13d
 
 runs:
   using: composite

AGENTS.md

@@ -91,11 +91,14 @@ nostr layer is enabled) — the same road every release ships.
   probe came back empty while the system journal carried the line.
   The failure dump's `journalctl -b` shape is the proven reader.
 - Displayless (`virtio-vga` + `QEMU_DISPLAY: none`), niri dies on
-  early import and the GPUI shell follows — "window not found" —
-  restart-looping until the unit's start limit ends the waves. A probe
-  that needs a live shell process is a coin flip wearing a deadline:
-  the notification name and the sleep inhibitor exist only while a
-  wave is. Ask what shipped and what the journal recorded instead.
+  early import and the pre-fix shell followed — "window not found" —
+  restart-looping until the unit's start limit ended the waves: a
+  probe needing a live shell process was a coin flip wearing a
+  deadline. kumaui's idle-without-windows fix (Oct 2026, e90884c)
+  keeps the shell alive holding its names and inhibitor through
+  output loss, and the notification and inhibitor probes are armed
+  again. The lesson stands: a probe that needs a live process asks
+  the runner's timing unless the process is guaranteed alive.
 - `egl-headless` wants a host DRM node a runner lacks. The GPU-less
   console combo is `virtio-vga` plus `none`.
 

notes/44.4.0-plan.md

@@ -426,10 +426,14 @@ binaries could not carry it. What the release week found:
   string. [coreos/bootupd#1166](https://github.com/coreos/bootupd/issues/1166) —
   any pipeline that composes label-less hits this on every boot.
 - **The smoke's real-session section learned what a displayless VM can
-  answer.** The shell survives the GPU-less VM in restart-waves until
-  the unit's start limit ends them, so every probe that needed a live
+  answer.** The shell survived the GPU-less VM in restart-waves until
+  the unit's start limit ended them, so every probe that needed a live
   shell process — notification ownership, the sleep-inhibitor
   readback, a stale noctalia pgrep — either measured the runner's
-  timing or passed vacuously. What remains asks the wiring and the
+  timing or passed vacuously. What remained asked the wiring and the
   journal; whether the shell draws is real hardware's question. The
-  measured facts are in AGENTS.md.
+  measured facts are in AGENTS.md. Re-armed 2026-10-03: kumaui's
+  idle-without-windows fix (e90884c) keeps the shell alive holding its
+  names and inhibitor through output loss, so the ownership and
+  inhibitor probes came back, and the kumaui pin was bumped to carry
+  the fix.

scripts/smoke.sh

@@ -1435,25 +1435,28 @@ smoke_published() {
             ok "the session started the shell's unit; whether it draws is real hardware's question"
 
             # Notifications: mako left with the swap, and the shell took
-            # the name. Whether the name is HELD is this VM's timing, not
-            # the machine's: displayless, the shell survives in
-            # restart-waves until the start limit ends them, and the name
-            # exists only while a wave is — a polled busctl here is a
-            # coin flip wearing a deadline. What CI asserts is that the
-            # daemon shipped (the binary and the unit, above) and that
-            # the session started it; whether a notify-send lands
-            # somewhere real is the same question as whether it draws,
-            # and real hardware answers both.
+            # the name. Since kumaui's idle-without-windows fix the shell
+            # holds the name for the session's life, displayless or not;
+            # polled anyway, because the unit starts seconds after
+            # loginctl can already see the session.
+            local owner_call='busctl --user call org.freedesktop.DBus'
+            owner_call="$owner_call /org/freedesktop/DBus org.freedesktop.DBus"
+            owner_call="$owner_call GetNameOwner s org.freedesktop.Notifications"
+            local notif_deadline=$((SECONDS + 60))
+            until guest "XDG_RUNTIME_DIR=/run/user/\$(id -u) $owner_call" >/dev/null; do
+                [ $SECONDS -lt $notif_deadline ] || bad "nothing owns org.freedesktop.Notifications in a live session"
+                sleep 5
+            done
+            ok "the shell owns org.freedesktop.Notifications"
 
             # Lock before suspend: `lock_before_suspend = true` is baked,
             # and the readback is the shell's logind sleep inhibitor.
-            # That readback is a real-hardware question in this VM — the
-            # shell survives here in restart-waves, and an inhibitor
-            # exists only while a wave is, so asserting it asserts the
-            # runner's timing rather than the machine's. The wiring that
-            # makes it true is kumaui's to hold; the guard below proves
-            # the no-shell half of the property, which is the half this
-            # VM can see.
+            # The same kumaui fix keeps it held through output loss; the
+            # guard below still proves the no-shell half of the property.
+            guest systemd-inhibit --list --no-pager \
+                | awk '$1 == "kuma-shell" && $6 ~ /sleep/ { found = 1 } END { exit !found }' \
+                || bad "the shell holds no sleep inhibitor: this machine suspends without locking"
+            ok "the shell inhibits sleep to lock first"
 
             # And the guard for when the shell is not there at all.
             #