[tannevaled]

Found while working out how to land the Ruby 4.0 version bump for #799, which
needs BASELINE re-recorded. It turns out nothing was watching that re-record.

-update never read the baseline it replaced

The entire body was:

f, err := os.Create(*baseline)
write(f, got, header)

So the one thing -update could never report is the one thing that matters:
whether the sweep replacing the record was better or worse than the record.
The comparison simply never happened.

It was written down twice, and gated in neither place

scripts/conformance/rubyspec/cmd/guards/main.go:18
    "UPDATE_BASELINE=1 records whatever it sees and would absorb a loss"

.github/workflows/rubyspec-ratchet.yml:41
    "the same run also showed kernel/require +1, equally environmental,
     and UPDATE_BASELINE=1 would bank it as progress."

Two notes about the same gap, written months apart, neither able to stop
anything. The second one is from the incident that pinned this lane to
ubuntu-24.04 — setrlimit read 28 on one Ubuntu major and 29 on the next,
same commit. The lane is pinned precisely because the environment changes the
answer, which is exactly the condition under which a blind re-record is at its
most dangerous: it would bank the image roll as progress and erase the evidence
in the same write.

The refusal

You re-record to lock in a win. A file that went backwards is either a real
regression or an environment that moved, and both need a person.

$ UPDATE_BASELINE=1 scripts/conformance/rubyspec/run.sh
re-record: 2206 files measured, 2206 in the baseline being replaced

-- REGRESSED (1) --
   core/a_spec.rb: 9 passing, baseline 10 (-1)

::error::REFUSING to re-record: 1 file(s) went backwards -- named above
A re-record locks in a win. Fix the regression, or pass -absorb to record
it on purpose -- the baseline then carries a line naming every file absorbed.

errWouldAbsorb is kept distinct from errRegression in the source though
both exit 1: "this branch broke something" and "I declined to erase the
evidence that something is broken" send a reader to different places.

The derogation stamps the artefact, not the log

A console warning scrolls away; a baseline outlives the run that wrote it by
months. -absorb writes the loss into the file:

# ⛔ ABSORBED 1 file(s) that went backwards, recorded with -absorb on 2026-11-03:
#     core/a_spec.rb: 9 passing, baseline 10 (-1)

Controls, beside each refusal

case	why it is here
a refused re-record leaves BASELINE byte for byte as it was	a refusal that truncated the file would destroy the record it declined to replace
a re-record that only improves is accepted, unstamped	without it, the rest only proves the gate always fires
a first recording, with no baseline to be worse than	an absent baseline is not a loss
a file that stops loading, and one that vanishes from the sweep	the other two ways backwards, same refusal

The test fixture's loss is a_spec 10 → 9 while b_spec gains — a net of
zero, which is the shape no total can see. The existing round-trip test already
re-recorded an improvement, so it passes unchanged: the cheapest control of all.

Refs #799

🤖 Generated with Claude Code