The scripts assumed our own split setup: a remote Ubuntu builder reached
through the "lobevm" ssh alias, plus a Windows QEMU install at a hardcoded
path. Nothing in the repo told anyone else what "lobevm" was, so on a fresh
clone not one script ran. Build locally, look for QEMU on PATH, and record
in the README the toolchain versions we actually tested against.

The two architectures cannot share one build tree: the Makefile writes .o
files next to the sources and its clean target deletes kernel/kernel and
fs.img, so whichever build ran second reused the first one's objects. The
remote setup hid this behind two separate directories. build-loong.sh now
stages a copy under build-loong/ and cleans it before building, since the
copy otherwise drags in riscv objects that loongarch ld rejects.

Add checkenv.sh to report toolchain versions and name the apt packages for
whatever is missing, and fsck-images.sh to run the e2fsck / fsck.fat checks
the README claims. Track run-disk.sh, which the README already referenced,
and drop the README's pointers to USAGE.md and patches/ -- both gitignored,
so they were dangling for anyone cloning. Documentation moves from a netdisk
link to docs/.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>