The federal layer never shipped --space-*, --radius-* or --z-* — so ten members invented token names and let the fallback carry the value

Why Care?

The first convergence sweep went looking for duplicated components. It found them. But all four scanners independently surfaced the same deeper thing, and it reframes what the duplication is:

--space-*    0 declarations in packages/theme/theme.css
--radius-*   0 declarations
--z-*        0 declarations

Members needed spacing, radii and layering. The federal vocabulary did not have them. So members invented the names they expected to exist and let CSS's fallback argument carry the real value:

170 declarations, 10 of 19 members. Every one resolves to its hardcoded fallback, identically, in all three modes. They do not respond to the mode switch. They cannot be re-skinned. They are invisible to the portal, to design-drift.mjs, and to every agent that reads token names rather than computed values.

This is not a discipline failure

It matters that the diagnosis is right, because the fix follows from it.

Nobody ignored the contract. Members reached for a vocabulary that the federal layer advertised the shape of and never delivered. Three members' own file headers assert that --space-* and --radius-* "come from @augment-it/theme" — they were written by people who believed the scale existed.

The consequences are visible everywhere the sweep looked:

• Four different radii for one control. 5px, 4px, 3px, 999px for the same chip across members; 8px / 6px / 4px for the same card.

• Five different paddings for one button. 5px 10px, 0.6rem 1.2rem, 0.3rem 0.7rem, 0.55rem 1rem, 0.25rem 0.6rem.

• 158 button rule-sets and 34 badge treatments federation-wide, none of them components. response-reviewer alone holds 16 ways to draw a button and 12 distinct pill recipes.

• sort-filter-lens renders magenta-bordered chips on a blue fill in every mode, because var(--color-accent) resolves to the federal electric-magenta while var(--color-accent-bg, rgba(120,160,255,.12)) falls back to a blue literal the theme never chose.

• Nine members render in a fallback font for anything reading --font-sans, in a product whose stated identity is "a dense monospace instrument panel."

Promoting a component before shipping the scale would bake the drift in. A promoted Button has to choose one radius and one padding. Choosing them from literals — rather than from a named scale — makes one member's accident into federal law. The token families come first.

F4 is currently unsatisfiable

F4 forbids raw z-index integers and requires the federal --z-* tokens. There are zero --z-* tokens. The rule cannot be complied with by anyone.

It is enforced in two places — scripts/design-drift.mjs and packages/gallery/src/audit.ts — so the federation ships a check that every member must fail. The shell itself carries 5 z-index literals; corpora-curator's gallery catalog pre-emptively confesses to one.

A rule that cannot be obeyed teaches people to ignore the checker. That is a worse outcome than not having the rule, and it puts every other F-rule finding at risk of being skimmed past.

--focus-ring is declared and almost unused

Declared in all three mode blocks. Consumed by 3 of 19 members (shell, docs-portal, corpora-curator) plus packages/gallery.

packages/theme/theme.css has no global :focus-visible rule outside its @media (forced-colors: active) block, so every member pays for its own focus styling and almost none do. Neither packages/shared-ui component defines one — the package that is supposed to model the contract fails the gallery's own Focus check.

One federal declaration fixes this in every member simultaneously:

*:focus-visible { box-shadow: var(--focus-ring); outline: none; }

The proposed work, in dependency order

Phase 1 — ship the missing families. --space-*, --radius-*, --z-* in packages/theme/theme.css, derived from what members actually use today rather than from a theoretical scale. The measured values are the requirements document: radii cluster at 3/4/5/6/8/999, paddings at the five listed above. Round to a scale that covers them; do not invent an eight-step ramp nobody asked for.

Phase 2 — name the real tokens the phantoms were reaching for. --color-warn (48 uses) is the biggest: the federal names are --color-warn-bg / --color-warn-text, and members wanted a single foreground. Decide whether to add the alias or migrate the 48. --font-sans (27 uses, 9 members) needs a decision about whether a sans face exists in this product at all — if it does not, those 27 declarations are a bug in nine members, not a missing token.

Phase 3 — add the global :focus-visible. One declaration, whole-federation effect, and it unblocks the a11y floor F7 already requires.

Phase 4 — migrate the 170 declarations, member by member, each as its own commit. Mechanical once the scale exists.

Phase 5 — only then revisit ButtonRecipe and StatusBadge as promotion candidates. They are the two largest duplication surfaces in the product and both are blocked on Phase 1, not on anyone's agreement.

The second-order defect — discovered 2026-09-13, when the scales shipped

Shipping the token families fixed the first-order problem: 170 declarations that resolved to a hardcoded fallback now resolve to a token. But it exposed a second one that was invisible while the tokens were missing, and this one is more interesting.

While a token is undefined, the fallback is the real declaration and the name is decorative. An author writing var(--radius-md, 6px) was shipping 6px; --radius-md was a label nobody could check, so they picked whichever name sat next to the pixel value they wanted. The moment the token ships, the name becomes load-bearing — and every mismatch surfaces at once.

apps/chat is the worked example, found during Phase 1's visual check. All eleven of its radius declarations sit one scale step below the role DESIGN.md §Shapes assigns them: seven --radius-sm uses are a small card, a panel, an input and four buttons — every one of them --radius-md work — and its four --radius-md uses are a bubble, a composer, a send button and a popover, all --radius-lg work.

The correction is pixel-neutral. sm→md restores 4px; md→lg restores 8px. Chat renders identically and its vocabulary becomes true.

Why this matters beyond chat

All 22 radius declarations in the product carried a fallback — there is not a single bare var(--radius-*) anywhere. So the conditions that produced chat's mismatch existed in every member that referenced a phantom token, across all six phantom families and all 170 declarations. Chat is not special; it is the first member anyone checked.

That reframes the remaining migration work:

• A member adopting the scales is not a find-and-replace. Every var(--token, literal) needs its role checked against the token's documented meaning, not just its pixel delta measured.

• The pixel delta is the wrong diff to review. Chat's most alarming change (4px → 2px, a visible flattening) was not the scale being wrong — it was the scale correctly revealing a wrong name. Reviewing only "did it move" would have concluded the scale was too tight and softened it, entrenching the bug.

• Some corrections will be pixel-neutral like chat's. Some will not, and those are the ones worth arguing about.

What to do about it

Nothing federation-wide yet, deliberately. Each member's radius roles get checked when that member is migrated, by the engineer already reading its CSS — which is the cheapest possible moment and the only one where the context is loaded. A pre-emptive sweep would produce a list nobody is positioned to act on.

Recorded here so that migration engineers are told to look, and so the next reader does not repeat Phase 1's near-miss: reading the pixel change and concluding the scale was wrong.

A check that would have caught this

Phantom tokens are mechanically detectable: every var(--x) in a member, minus every --x: declared in packages/theme, is the phantom set. That is a ~20-line addition to design-drift.mjs and it would have fired on the first of these 170 declarations rather than the hundred-and-seventieth.

It is the same lesson as Structural Invariants Live In Prose So Sweeps Stop Halfway: the contract asserted a vocabulary, nothing compared the assertion to the runtime, and the gap grew silently for months. design-drift.mjs compares CSS to CSS; nothing compares what members consume against what the theme provides.

Related

• Design System Convergence — the organ registry; several verdicts are blocked on Phase 1

• Converge The Federated Design System — the sweep that surfaced this

• The Sort Filter Lens Containment Breach — the other federal defect from the same sweep

• Structural Invariants Live In Prose So Sweeps Stop Halfway — same disease, build config

• DESIGN.md §Token architecture · §The federation contract (F4, F7, F8, F11)