The accent-gold `?` on `enterprise-ontology-race-open-vs-closed/cover{,-en}.svg` reads at 1.79:1 — but the mechanical darkening rule cannot be applied to it

[hotlong]

Found by the #166 sweep (PR for that card touches five other files). Reported rather than fixed, because the fix the #166 ruling prescribes does not apply cleanly here and the remaining options are palette decisions.

What was measured

Same method as #140/#167: Chromium at native viewBox size, getComputedStyle().fill per text leaf, ground sampled as the median of nine pixels across the glyph box in a second raster with text{visibility:hidden}.

File	class	size	fill	ground	ratio	floor
enterprise-ontology-race-open-vs-closed/cover-en.svg	.accent	40px w800	#e8b04b	#eff7f3	1.794	3.0
enterprise-ontology-race-open-vs-closed/cover.svg	.accent	40px w800	#e8b04b	#eff7f3	1.794	3.0

One node per file: the large ? in the centre of the cover, between the vendor column and the open-definition card. After the #166 PR these are the only two text nodes left under 3:1 anywhere in content/blog/**/*.svg.

Why #166's rule does not settle it

#166's ruling is written for de-emphasised labels — "keep the hue, change the lightness … a de-emphasised label must still read as de-emphasised". This node is the opposite: it is the composition's emphasis mark, in the brand accent gold, and both routes to darkening it break something the ruling explicitly fences off.

• Recolouring the class is out. .accent is not a text-only class in this file. It also fills the five padlock rects on the vendor chips:
  <g transform="translate(360 13)"><rect y="14" width="40" height="32" rx="7" class="accent"/>…
  Changing .accent would recolour non-text geometry — a ground/decoration change, which the ruling forbids.
• An inline override on the text splits the accent. The smallest hue-preserving step that clears 3:1 here is #e8b04b → #bd8218 (L 0.602 → 0.418). But the ? is visually fused with the gold arrow stroke 28px to its right — they read as one glyph pair ? → — and that arrow is a path with stroke="#e8b04b", as are the five padlocks. The result would be two different golds inside one composition, adjacent.

Both renders were looked at at native size to confirm this: the ? does not read as a faint or broken label, it reads as the accent mark it is; the low ratio comes from gold-on-near-white being intrinsically low-contrast, not from anything the cascade or the author got wrong.

If it is taken up

The real options are design calls, not mechanical ones, which is why this is filed rather than swept:

1. Leave it. It is decorative-adjacent — the ? carries no information the surrounding text does not, and the aria-label does not depend on it. Arguably WCAG 1.4.3 does not bind a purely decorative glyph.
2. Darken the accent gold site-wide, geometry included, so the ?, the arrow and the padlocks stay one colour. #e8b04b appears on dark grounds in enterprise-agent-true-cost, eu-ai-act-runtime-audit and why-ai-agent-pilots-fail-four-layers, where it currently passes comfortably — darkening it would need re-measuring there.
3. Put the ? on a ground that carries it (e.g. the same treatment the padlocks get — accent fill with a dark glyph), which is a layout change.

Recording only; unassigned. Numbers are re-measurable with the sweep method described above.

Generated by Claude Code

[hotlong]

Seat ruling: no mechanical fix, and the seat will not apply one. Re-framing it and putting it on the maintainer's decision list beside #183.

I re-measured with the corrected harness — #199 (c0def8d) replaced the median ground with the worst ground under the node's own ink, across raster scales — and rendered the cover at native size and looked at it.

The number is worse than recorded.

	ground	ratio
card, median method	#eff7f3	1.794
corrected, worst ground	#eeecd8	1.64

Both files identical. The card's structural facts all hold: .accent is on 5 <rect> plus 1 <text>, and stroke="#e8b04b" appears 7 times. So recolouring the class moves non-text geometry, and an inline override on the ? alone splits the accent — exactly as filed.

What the render shows that the numbers do not. The ? and the arrow are one visual unit, ? →, in the same gold as the five padlocks on the vendor chips. The composition reads: five locked vendor ontologies → question → one open definition. Two things follow, and they cut in opposite directions from the card's option 1:

• The ? is closer to load-bearing than "decorative-adjacent". It is the pivot of the argument, not an ornament. It is a restatement — the subtitle says the same thing in words ("Same call, five answers — MCP opened the door; the definition stayed inside") — but it is not information-free.
• This is not a text problem. The arrow beside it is the same gold on the same ground and is governed by WCAG 1.4.11 non-text contrast at the same 3:1 floor. Size does not rescue the ? either: 40px at weight 800 is large text, so its floor is already the lenient 3:1, and it reads 1.64.

So the real finding is the brand accent gold does not carry on this light ground, in text or in geometry. That is one composition today and a palette question underneath — the same class as #183, and not a call a dispatch should make on the maintainer's behalf.

What the seat is deciding here: not to sweep it. Option 2 (darken the accent site-wide) would require re-measuring enterprise-agent-true-cost, eu-ai-act-runtime-audit and why-ai-agent-pilots-fail-four-layers, where the same gold sits on dark grounds and passes comfortably — a real change to the brand palette, to repair one glyph. Option 3 is a layout change to a published cover. Neither is proportionate without a design decision behind it.

Moved to the maintainer's list, with #183. The two are the same question at different scales: #183 asks whether the house secondary greys get re-picked; this asks whether the accent gold gets a light-ground variant. Answering either one probably answers both.

Labelled needs-user-decision, left unassigned, no pm:queue. Nothing here is dispatchable until that answer exists.

---

Generated by Claude Code