A freshness indicator that measures a different object than the one you compared turns "nobody looked" into confidence

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

Do not derive a “last updated” line from one object and then compare a different one. porthole pkg drift was written to catch a carried fork being overtaken upstream. It read each upstream APKBUILD from the checked-out working tree of aports_upstream, and printed its freshness line from the mtime of .git/FETCH_HEAD. Those are two different objects, and a plain git fetch moves only the second:

worktree HEAD           2026-08-20   ## master...origin/master [behind 1009]
.git/FETCH_HEAD mtime   2026-09-09   <- what the tool printed
main/mesa (worktree)    pkgver=26.1.6
main/mesa (origin/master) pkgver=26.2.2

So the tool printed

upstream tree last fetched   2026-09-09
  mesa   26.1.6-r14   upstream 26.1.6-r0   SAFE

while mesa’s three a5xx patches were in fact already outranked. The verdict was not merely wrong; it was wrong with a timestamp asserting it was current, which is worse than no tool at all. A missing check invites you to look. A confident all-clear stops you looking.

Why this is easy to get wrong. git fetch feels like “update the tree”, and in a checkout nobody edits it is easy to believe the working files track the remote. They do not: fetch updates remote-tracking refs and leaves the worktree exactly where it was. Any tool that reads files from a fetched-but-not -merged checkout is reading history.

What to do instead. Read the ref you are claiming freshness for, and date the ref you actually read:

git -C <repo> show <ref>:<path>            # not the worktree file
git -C <repo> log -1 --format=%cs <ref>    # date of what you just read

Reading a ref is offline, so this costs nothing and adds no network side effect. If you keep a worktree fallback for checkouts with no remote, make the output say which source each row came from – a silent fallback to the stale thing is the original bug wearing a different hat. And say so when the worktree itself is behind, because the reader will otherwise draw conclusions about a checkout they think is current.

The general shape, beyond git. Whenever output pairs a value with a statement about that value’s age or provenance, check that both come from the same object. A cache’s timestamp beside a recomputed value, a package index’s date beside a version read from disk, a probe’s “as of” beside a number from a different read – the same trap, and it always resolves the same way: date the thing you read, or read the thing you dated.

Related: the-instrument-is-guilty-until-proven-innocent, an-instrument-that-fails-quietly-is-worse-than-none.