The default kernel tree can be parked on an old branch, and then every tool reports the wrong kernel version perfectly truthfully

scope: generic · severity: trap · confidence: proven · subsystem: build

Evidence

google-taimen, 2026-09-19. The device had run 7.2.2 since aport r52 and was on #60/r59 that day, but /linux was parked on the 6.18 topic branch wifi-disablekey-test with 19 uncommitted files. porthole build --measure correctly printed ‘this tree is Linux 6.18, the device runs 7.2.2’ and refused; porthole doctor’s kernel-tree row mentioned only shallowness and never the version. The 6.18-vs-7.2 confusion was re-derived across multiple sessions from exactly that, and the owner had to raise it.

The trap — a kernel tree is a working copy, not a statement about the port. PORTHOLE_KERNEL_TREE defaults to <workdir>/linux, and nothing keeps that checkout on the product branch. Leave it on a topic branch from two kernel versions ago – which is the normal end state of any bisect or feature branch – and every tool downstream reports that version, accurately, forever.

The failure is not that a tool lies. It is that “the tree is 6.18” and “the port is 6.18” are different statements that look identical in output, and a reader has no way to tell them apart from the line alone.

What it looks like

>> NOTE: profile says the product branch is taimen-v7.2
>>       this tree is on wifi-disablekey-test -- one of the two is stale
>> WARNING: this tree is Linux 6.18, the device runs 7.2.2

That NOTE is the tell, and it only appears once you try to build. Anyone who read doctor or brief first never saw it.

How to not fall in

• The device is the authority on the device: cat /proc/version. The aport name is the authority on what ships (PORTHOLE_KERNEL_PKG). A tree is neither.
• Check a branch’s version without checking it out: git show <branch>:Makefile | head -3.
• KBUILD_BUILD_VERSION = pkgrel + 1, so #60 is aport r59. Reading the #NN as the pkgrel is off by one every time.

Fixed, in three places, because one was not enough

1. PORTHOLE_KERNEL_TREE pinned to a worktree that stays on the product branch (git worktree add ../linux-<series> <product-branch>), in config.env and systemctl --user set-environment – the user manager holds a second, invisible copy of the config.
2. doctor’s kernel-tree row now compares the tree’s Makefile version against the series in PORTHOLE_KERNEL_PKG and warns before it says anything else, naming both and how to fix it.
3. The stale checkout was left exactly as found. It was carrying 19 uncommitted files that were not mine to discard, and adding a worktree costs a checkout while git checkout over someone’s WIP costs their day.