SkillFishOS 26.06.4 "Aetherium" is out

[MTSistemi]

Two editions — BC-250 and Generic x86-64 (PCs and virtual machines) — from
skillfishos.com/download, over SourceForge or BitTorrent.

cad444adccae6a4da757171894839b56e01878e8630972966cbe389e029cb261  SkillFishOS-26.06.4-Aetherium-BC250-amd64.iso
235fde126655f476028c00b6fecfcf8603891c4f61c0ab7706138fd08d679fa7  SkillFishOS-26.06.4-Aetherium-Generic-amd64.iso

Two defects that could stop an installation

The installer could reboot the board out from under you. SkillFishOS unlocks the two
CPU cores the BC-250 keeps disabled, and the only way to do that is to write the SMU core
mask and reboot — that is how the board re-reads how many cores it has. Nothing stopped it
from doing that inside the live session. The board rebooted a few seconds after start-up
and, if that landed while Calamares was copying, the installation died. The unlock is now
off by default and opt-in from the Tuner.

We never caught it because every install test ran in a virtual machine, where the SMU
write fails, the script exits and never reboots. The code path that breaks real hardware
is exactly the one a VM cannot execute. That is the lesson from this release, and it is
now a roadmap commitment: every image gets installed on a real board before it ships.

Installing onto Btrfs could produce a system that would not boot, with
error: premature end of file /@/boot/vmlinuz-… at the GRUB prompt. Calamares copies with
rsync -aHAXS — that S is --sparse — and our kernel image ends in 512 zero bytes, so
it landed with a hole at the tail. On Btrfs a hole has no extent, GRUB walks the extent
map, sees it end before the declared size and stops. Linux reads the file perfectly, which
is why nothing looked wrong. ext4 was never affected.

Worth saying plainly: we had already shipped a "fix" for this in 26.06.3. It ran, logged
that it had rewritten two files, and left the defect byte for byte identical, because it
used cp --reflink=never without --sparse=never. The CI guard meant to catch the
regression searched for a string present in the broken version too. Code, changelog and CI
agreed with each other for six days.

Also fixed

• The disk check asked for 10 GiB when the system needs 15, so a small disk passed and
  then failed part-way.
• The BC-250 edition installed, then booted the generic kernel.
• Safe Mode and Text Mode in the boot menu did exactly the same thing as the normal entry.

Earlier 26.06.x images have been superseded — if you have one, take 26.06.4.

Full notes: RELEASE-NOTES-26.06.4.md