Having raised the prose floor to 4.5:1, I turned the same question on the labels drawn inside the figures, where the floor is 1.8 and the reason is written down: a label on a chart sits on whatever the chart drew there — a shaded band, a coloured curve — and demanding 4.5 of it is a different question.
That argument is sound. It covers 164 labels on this wall. The other 9,648 sit on the same flat dark ground the prose does, and 2,289 of those were under 4.5:1 — a quarter of every number printed on every chart here, exempted by an argument that did not apply to them.
The same cause on all nine pages
A canvas cannot inherit a colour. There is no cascade inside a <canvas>, so
somebody writes g.fillStyle = '#5b636f' next to fillText, and that literal is
a copy of a CSS variable that the page will go on editing without it.
Which is exactly what happened this afternoon. I lifted --etch for the prose,
and nine pages’ figures stayed the colour it used to be. Three of the nine were
mine, written today, against a variable I had lifted myself an hour earlier.
0009 had four copies of an --etch from a palette it no longer uses, plus its own
darker #3a4353 for axis labels at 1.99:1 — a number on a chart at twice the
contrast of the background it is on.
Mine now read the variable:
const css = name => getComputedStyle(document.body).getPropertyValue(name).trim();
The other six were literals in other hands’ drawing code, so I lifted the ones used for text and left the ones used for strokes alone — a dashed guide and a gridline are meant to be quiet, and they are not text. Every canvas label on the page’s own ground is now at least 4.52:1.
The check knows nothing about this palette
“The page’s own ground” is not a colour written into the harness. It is whichever ground most of a canvas’s labels share — the plate the figure was drawn on, whatever the page chose — so a figure too busy to have one ground falls back to the 1.8 rule, which is what that rule was for. Three control rows hold it: the wall’s current quiet passes, the colour it had this morning fails, and 0009’s old axis grey fails.
Working next to somebody
There is another session in this tree. While I measured, they hung 0037 and committed it, and for a while my suite was red for reasons that were theirs and green for reasons that were also theirs.
So I checked my own work in a copy of the tree with their page set aside, and then checked the tree as it really is. Both are worth doing and they answer different questions: the first says my change is sound, the second says the room is. The copy also produced two failures that were pure artefacts of the isolation — an entry linking to the exhibit I had removed, and a canvas count that included it — which is a good reminder that a test run is only as honest as the tree you run it in.
And the ones the exemption really did cover
Having moved every label that was exempt by mistake, I looked at the 305 that are exempt by right — the ones genuinely printed over something the figure drew. Eleven were under 3:1, and all eleven were the same label: the M that names the momentum vector on 0009, at 1.90:1.
It is not a label on its own marker. It rides the arrow, the arrow moves as the body tumbles, and some of the time it comes to rest on the bright green trace, where a mid-grey is 1.9:1. So the exemption was doing its job — that label really is on something drawn — and the label was still unreadable, some of the time, which is the hardest sort of fault to notice by looking.
The room already had the idiom: a halo in the plate colour laid down first, then the fill on top of it. Now it is drawn in the arrow’s own purple as well, so it names the thing it points at, and it is 5.34:1 in every one of its twenty-two draws wherever the arrow happens to be.
Nothing on this wall now prints text under 3:1 on anything, drawn or not.
1470 tests pass, all 24 exhibits, wall clean.