image: the serial getty is Type=simple, which gives back v20261113.01's console prompt

boot/bench_test.go

@@ -87,6 +87,9 @@ func without(units ...string) variant {
 //   - tmpfs /tmp against /tmp on the disk: 3 ms, within noise (220/232 against 223/230, 20
 //     boots, 2026-09-26); image/build.sh forbids the disk.
 //   - serial-getty Type=simple instead of Type=idle: within noise of the echo (2026-09-10).
+//     And again once spin lingered and user@1000.service was in the boot transaction: 346/358
+//     against 423/441, 20 boots, lab run 43362067503 (2026-11-13) - now the image's
+//     (serial-getty@.service.d/type.conf).
 //   - loop.max_loop=0 8250.nr_uarts=1, and CONFIG_VT=n: no change to a login prompt
 //     (2026-09-26). udev's coldplug overlaps the boot rather than delaying it.
 //   - 19 udev rule files for hardware the machine cannot have, masked: no change, same day.

image/build.sh

@@ -299,6 +299,13 @@ for f in /etc/systemd/system/serial-getty@.service.d/term.conf \
     in_image "$f" || { echo "ERROR: the image has no $f; a login would inherit TERM=dumb" >&2; exit 1; }
 done
 
+# The serial getty is started Type=simple: Type=idle holds it until the boot's job queue
+# drains, which with spin lingering is 77 ms of a login prompt (serial-getty@.service.d/type.conf).
+"$bin/debugfs" -R "cat /etc/systemd/system/serial-getty@.service.d/type.conf" /work/base.raw 2>/dev/null |
+        grep -qx 'Type=simple' || {
+    echo "ERROR: the serial getty is not Type=simple; its prompt waits for the job queue to drain" >&2
+    exit 1; }
+
 # A boot optimization that is invisible in a file listing and expensive when it regresses:
 # a masked unit is a symlink to /dev/null, and a package upgrade replacing one turns it
 # back on. One of the ones optimize-systemd.sh masked for measured reasons.

image/mkosi.extra/usr/local/lib/spin-base/rootfs/etc/systemd/system/serial-getty@.service.d/type.conf

@@ -0,0 +1,14 @@
+# The serial login prompt as soon as its getty is started, not once the boot's job queue drains.
+#
+# The template is Type=idle: systemd forks the getty and holds it until no job is left running.
+# That cost nothing while the queue was empty by then, and Type=simple measured within noise of
+# it (2026-09-10). Since spin lingers, logind starts user@1000.service and its chain inside the
+# boot transaction, and the getty waited for them: 91.8 ms between its fork and its exec in the
+# median boot of v20261113.01 (task boot:waits, lab run 43361486792), against 3-5 before.
+# Type=simple gives the prompt 77 ms back - 346/358 against 423/441 ms p50/p95 to a login
+# prompt, 20 boots interleaved on the bench fleet (task boot:bench, lab run 43362067503).
+#
+# What Type=idle is for is a prompt not overwritten by the boot's own status lines, and this
+# machine's console prints none: the kernel is quiet and systemd shows no status.
+[Service]
+Type=simple