Hi Junhui,

Attached are the exact minimal board DTS and SoC DTSI used for both runs,
plus the configuration extracted from the tested Image and reproduction
notes. I rebuilt the DTB from the recorded source and recovered config;
it matches the tested 4497-byte DTB byte-for-byte (SHA-256):

  4c7b762757392334e8b7ac525e60bf5757ec23b34efdcc949d2093c238b1ed16

The support archive contains the two local integration/comparator diffs,
the four A733 clock/reset headers, source history and short UART excerpts.
Applying the 19 published patches and these local diffs to v7.3-rc1
reproduces the tested source tree. This is the minimal diagnostic DT,
not my newer DTS work; its fixed card supply is part of that test setup.

Both the September 21 failing run and September 26 clk_ignore_unused run
report identical firmware banners:

  U-Boot 2018.07-12-boot-aw2501-gb2d229198b2-dirty
  (Jan 06 2026 - 03:45:33 +0000) Allwinner Technology
  BL31: v2.5(debug):5fc237a6a
  BL31: Built : 09:05:28, Feb 26 2025

These tests used the vendor firmware already installed on the A7S;
I did not build or flash U-Boot or TF-A for them. Unfortunately, I cannot
establish the original installation/download image, the U-Boot dirty
delta, or the BL31 source/build inputs. I also do not have verified hashes
of the installed firmware binaries. Earlier source-correlation records
point to Radxa's downstream U-Boot, but the retained candidate build is
a different revision and is not a reproducible source for this firmware.

reproduce.txt includes the actual ext4load/booti commands and bootargs.
In particular, I set the U-Boot environment variable drm_debug=1 as well
as including it in bootargs. U-Boot reports FDT fixup errors and "update
dts" before Linux, also on the successful run; the excerpts retain those
messages. The DTB hash above describes the file loaded into RAM. I did
not capture the final post-fixup tree passed to Linux, so I cannot yet
rule out a firmware or DT-fixup difference from your A7A.

Assisted-by: Codex:gpt-6

Regards,
Enzo