Power-button standby (suspend-then-hibernate) was intermittently bouncing
back within seconds instead of staying asleep. journalctl showed
suspend/resume cycles as short as 3s correlating with Thunderbolt dock
(re-)enumeration and AC-adapter transitions, while undocked suspends ran
for hours as expected. This SoC only supports s2idle, which is far more
sensitive to spurious wakeups than deep sleep.
Disable ACPI wakeup on the Thunderbolt PCIe root ports (TRP0, TRP2) via a
oneshot service, and disable USB remote-wakeup on the dock's Realtek
Ethernet adapter (0bda:8153) via udev. Lid, power button, and USB input
device wakeup are untouched.