Two palette steps share one hex, so dark-mode borders measure 1.00:1

Why Care?

--color__graphite-800:  #232634;
--color__graphite-700:  #232634;

Two named steps of the Tier 1 palette, one value. In dark mode --color-surface resolves to graphite-700 and --color-border resolves to graphite-800, so a border drawn with --color-border on --color-surface measures exactly 1.00:1 — it is not a faint hairline, it is not there.

This is the failure mode the session has been naming all day, arrived at from the other direction. The rule was two hex declarations are two values and will drift. Here two names converged on one value, which is worse in a specific way: the semantic distinction between surface and border survives in the source, in the docs and in every review, and evaporates only at paint time.

It is dark-mode-only. Light measures 1.41:1 and vibrant 1.45:1, both inside the 1.2–1.5 band gate A22 documents as intended for a hairline. Dark is supposed to be in that band too and is not.

The four findings, adjudicated

A corpora-curator migration probe dumped resolved tokens off document.documentElement in both modes and produced four federal boundary claims. Three hold. One does not, and the one that does not is a mistake this session has now made twice, so it is worth recording as carefully as the real ones.

1 · CONFIRMED — the palette collision

Above. Fix is a Tier 1 edit: give graphite-700 a value distinct from graphite-800, landing dark's border-on-surface in the same 1.2–1.5 band light and vibrant already occupy. The palette's own ordering convention (higher number = darker) says graphite-700 should be the lighter of the two, so it moves, not the border.

Blast radius: every member, every surface with a border. Confidence: measured (and independently re-derived from the source hexes, not taken on report).

2 · REJECTED as stated — but it exposes a real narrower bug

The claim was that --color-border-strong fails F7's 3:1 control-boundary floor in dark mode at 2.73:1, and therefore Button's secondary and outline variants fail in dark while passing in light.

secondary does not fail. It sets background: var(--color-surface-raised), so its border sits on surface-raised, not on surface. Measured against the surface it actually uses:

The 2.73 figure comes from measuring against --color-surface. This is the identical error made earlier in the session — and it is an easy one, because the token is named for what it is, not for what it sits on.

outline is a different story, and here the claim lands. It sets background: transparent, so its border genuinely sits on whatever the parent painted:

So an outline button placed on --color-surface in dark mode is under the floor — and --color-surface is exactly where cards, panels and headers draw. The variant passes or fails depending on where a member puts it, which no component-level gate can catch and no member can be expected to reason about.

Fixing finding 1 changes this number, because graphite-700 is moving. Re-measure after, do not fix twice.

Blast radius: federation-wide, contextual. Confidence: measured.

3 · CONFIRMED — destructive has no control boundary at all

.ui-btn[data-variant='destructive'] {
  background: var(--color-error-bg);
  color: var(--color-error-fg);
}

No border-color, so it inherits transparent from the base. A filled control's boundary is its fill against the surrounding, and --color-error-bg measures 1.23:1 on the dark background, 1.02:1 on dark surface, 1.25:1 and 1.32:1 in light. Against a 3:1 floor.

The text inside is fine — 7.19:1 dark, 4.89:1 light — which is precisely why this survived nine migrations. Every contrast gate the federation runs is a text-contrast gate. A destructive button is a red-tinted patch with legible words on it and no perceivable edge, and the one control in the system where the edge matters most is the one that deletes things.

Blast radius: federation-wide, every destructive call site across fifteen members. Confidence: measured.

4 · CONFIRMED — ghost has no hover on a tinted surface

ghost's hover is --color-selected-tint, which aliases --color-accent-bg. Any chip, selected row or tinted panel is already painted with it. Composited, hover-against-rest is 1.122:1 — technically present, perceptually absent.

This is not hypothetical: corpora-curator's tag chips host a ghost icon button and are themselves --color-selected-tint.

Blast radius: federation-wide wherever a ghost control sits on a tinted host. Confidence: measured.

Why this is not being fixed right now

Four migration agents are in flight against org-workbench, response-reviewer, sort-filter-lens and shell as this is written. All four captured baseline gate numbers and are measuring before/after contrast in their own probes.

Editing packages/theme/theme.css mid-wave would move the federation number under them and silently invalidate every contrast measurement they have taken. That is the same hazard as the parallel-migration gate drift the playbook already warns about, except self-inflicted and worse, because a theme change is invisible in a member's own diff.

Queued behind wave 4. The order when it opens:

1. Fix finding 1 (Tier 1, graphite-700).

2. Re-measure finding 2 — the number changes because the surface moves.

3. Fix finding 3 (give destructive a border, or accept fill-only and say so explicitly in the contract).

4. Fix finding 4 (ghost hover needs a value that is not the tint it may be sitting on).

5. Re-run design:drift and the 30-pair contrast gate; expect movement.

The gate-shaped hole underneath all four

Every one of these is a control-boundary failure and the federation has no automated control-boundary check. The contrast gate walks 30 text pairs and passes 30 of 30 through every one of these findings.

F7 states a 3:1 boundary floor in prose. Nothing measures it. That is the same shape as Structural Invariants Live In Prose So Sweeps Stop Halfway, and the same remedy applies: a check that resolves each variant's boundary colour against each plausible parent surface, per mode, would have caught all four findings the day they landed rather than on the fifteenth migration.

Related

• Component API Contract And The Control Scale — F7 and the boundary floor

• Adopt The Shared Button In One Member — where the probe that found these is specified

• Structural Invariants Live In Prose So Sweeps Stop Halfway — the same shape, one layer up

• The Federal Layer Never Shipped Space Radius Or Z — the prior Tier 1 / Tier 2 repair