## Summary

On this machine, `systemctl hibernate` never powered off. The kernel aborted
**during image creation**, unwound, and half-resumed: display dead, keyboard
backlight lit, session frozen. Only a forced power-off recovered it.

Setting `HibernateMode=shutdown` fixes it completely.

## System

| | |
|---|---|
| Omarchy | 4.0.1-1 |
| Kernel | 7.1.9-arch1-2 |
| CPU | AMD Ryzen 9 8945HS w/ Radeon 780M |
| GPU | AMD HawkPoint1 (iGPU only, no dGPU) |
| Board | ASUS ROG Zephyrus G14 GA403UV |
| BIOS | GA403UV.311 (2025-06-09) |
| systemd | 261.2-1-arch |
| `/sys/power/mem_sleep` | `[s2idle]` — no `deep`, Modern Standby only |

Hibernation was configured by `omarchy-hibernation-setup`: 30.6G btrfs swapfile,
`resume=` + `resume_offset=` on the cmdline, zram at priority 100.

## What happens

With the default `platform` mode, `swsusp_save()` aborts **silently**:

```
PM: hibernation: Creating image
PM: hibernation: Need to copy 3457488 pages
PM: hibernation: Normal pages needed: 3457488 + 1024, available pages: 4744684
ACPI: PM: Restoring platform NVS memory      <- next line; already unwinding
```

`swsusp_save()` has three exits — `Image created (N pages copied)`,
`Not enough free memory`, `Memory allocation failed`. **None printed**, with
`pm_debug_messages=1`. Memory was not the constraint: ~1.29M spare pages (~5 GB).

It never reaches the image write or power-off, so the machine stays powered.
Reproduced twice, identically.

## Fix

```
# /etc/systemd/sleep.conf.d/hibernate-mode.conf
[Sleep]
HibernateMode=shutdown
```

Works first try, and the ACPI S4 bracketing disappears from the log:

```
PM: hibernation: Hibernation image restored successfully.
```

(That line is printed only on the resume-from-image path, so it confirms a real
write + power-off + restore cycle.)

## Scope — please judge this yourselves

**This is n=1.** One machine, one BIOS revision. I don't know whether it applies
to all GA403UVs, all ASUS boards, or all Modern-Standby AMD laptops, and forcing
`shutdown` globally could regress machines where `platform` works. You have the
fleet visibility to decide; I don't. A conditional — e.g. apply when
`/sys/power/mem_sleep` offers no `deep` — might be the safer shape.

## One observation

`default/systemd/system-sleep/keyboard-backlight` says it exists because
"the ASUS keyboard controller can block S4 shutdown if LEDs are active". On this
machine S4 never got far enough to be blocked, and hibernation now works with
that hook still not running (see #8588 — it is installed non-executable). That
may mean the hook is treating a symptom of this same firmware problem.

## Notes

Two pre-existing, non-blocking warts remain after the fix:

- `mt7921e ... PM: failed to restore async: error -110` — WiFi reloads firmware
  and recovers, so it blips on every resume
- `PM: freeze of devices complete after 5302.394 msecs` — 5-7s device freeze
  (`tpm_crb_acpi MSFT0101:00: Ignoring error -5 while suspending` is in the same
  phase)

Happy to gather more logs or test patches.