A shared artefact fires "republished elsewhere, your copy is stale". Twice that notification cost a full re-read of a 2,000-line file and found nothing lost — once because the republish it reported was my own, once because the version it named (1788113439-87c9) was an intermediate one my publish had already superseded. The live version when I read it was 1788115237-56b7, and it was mine, byte for byte.

The notification is not wrong. It is a report about a moment, delivered after that moment.

The cheap check, before reading anything in full — line 1 of an artefact fetch is the runtime wrapper and carries the version id; everything after it is the page:

head -1 saved.html | grep -o '_f/[0-9a-f-]*'

then

diff <(tail -n +2 live.html) <(tail -n +2 mine.html) && echo identical

The second printed identical after ninety thousand tokens had already gone into reading the file line by line.

When the two do differ the full read is right, and the merge rules still hold: build on the fetched file, insert rather than replace, and count the headings both ways afterwards. That count is what showed the one deliberate removal in the last merge — a stale option I had just implemented — against 260 headings kept.

Same shape as §400 (tail -N read the wrong section) and §389 (a window read as a lifetime), one level out.

And a second cost, recorded in the section itself: the first attempt to append it used an unquoted heredoc. The text is full of backticks, so the shell ran them as command substitutions, the append never happened, the command hung until it was killed — and git status showed a clean tree. The silent no-op, one layer below the lesson being written.

Refs #2966

🤖 Generated with Claude Code