[REGRESSION] NVIDIA dGPU never enters D3cold since Linux 7.3 merge window

Since upgrading past 7.2.4, my NVIDIA dGPU no longer entersD3cold/runtime suspend and stays permanently active. Confirmed
regression, not present on 7.2.4, present on both 7.3-rc1 and
7.3-rc2.

Hardware: ASUS ROG Zephyrus G16 (GU605MI), Intel Core Ultra 9
185H (Meteor Lake), NVIDIA RTX 4070 Max-Q / AD106M, nvidia-open
615.71.09 driver (unchanged across all tested kernels).

Last known good: 7.2.4
First known bad:  7.3-rc1

Evidence (/proc/driver/nvidia/gpus/0000:01:00.0/power):
  7.2.4:    Runtime D3 status: Enabled (fine-grained)
  7.3-rc1:  Runtime D3 status: Not supported
  7.3-rc2:  Runtime D3 status: Not supported

Same output from /sys/bus/pci/devices/0000:01:00.0/power/
runtime_status, power/control, power_state confirms the device
sits at active/auto/D0 permanently on the broken kernels, and
suspended/auto/D3cold on the working one. No processes hold
/dev/nvidia* open in the broken state, ruling out a userspace
handle as the cause.

I haven't been able to bisect further since the regression is
already present at rc1, meaning the change is somewhere in the
7.3 merge window. While investigating I came across an in-progress
ACPI series that touches related territory (PCI devices getting
stuck referencing power resources during early enumeration):
https://lore.kernel.org/linux-acpi/2295263.irdbgypaU6@rafael.j.wysocki/ (local)
but I want to be clear I haven't confirmed this is the same
mechanism, I checked dmesg for the specific _DEP/INT3472
trigger it describes and found no matches, so it may well be
unrelated. Mentioning it only in case it's a useful lead.

Note: tested on CachyOS's cachyos-rc kernel build (patches:
https://github.com/CachyOS/linux-cachyos), not vanilla. Happy
to test a vanilla build or apply patches if that would help
narrow this down.