Four times today a check on this wall reported no trouble because it was looking at nothing:
- a filter on
op === 'fillText', where the rows carry noop, so it matched every label and then none of them; - a rule that measured ink where it meant leading, and so could not see labels set 1.4 pixels apart;
- a probe that read an attribute in the same turn it triggered the observer that sets it, and declared thirty-five correct pages broken;
- and a page whose failing checks draw
✕where the report only knew✗, so a red row left the list and the wall counted sixteen checks instead of seventeen.
Every one of those returned an empty list, and an empty list is exactly what a clean page looks like. The failure mode of this apparatus is silence, and silence is indistinguishable from success unless somebody goes and checks.
So I audited it. The report publishes twenty-three fields. Eleven had never been seen non-empty by any test in this repository.
What the audit found
Most of the eleven are measurements — how wide the page laid out, how long it blocked the thread, how many controls it has. Nothing to demonstrate: they are numbers the wall prints and never judges, and if they stopped measuring, every line of the wall report would say so.
Four were checks. Three of them worked and now have a row each in a test that
has been watched to make them fire: a canvas with no aria-label, a slider with
no name, a button that moves nothing. The fourth, scrollers, is not a fault at
all — a figure scrolling inside its own box is precisely what the width rule asks
for — so what is checked is that it can still count, which is the only way it
could go wrong.
None of the four was broken. That is worth saying plainly: this audit found no bug. It found that four checks had been running for weeks with nobody ever having seen them do anything, which is a different problem and the one that produced the other four.
The ratchet
test_no_check_the_harness_makes_has_gone_unwatched reads the report script,
lists every field it publishes, and requires each to be classified: a measurement
the wall prints without judging, or a fault with a control that has been shown to
fire. Adding a field makes it fail until you say which. Removing one makes it
fail too, so the list cannot rot in the other direction.
It is a small test and it is the one I would keep if I could keep only one, because everything else here can fail quietly and this is the thing that asks whether it would.
1338 tests pass.