refrakt contracts is documented as describing "the complete HTML structure the identity transform produces for every rune". It does not. Measured across the checked-in contracts/structures.json (132 runes), not one universal axis appears:
Tracking started Sep 2 — check back for trends.
generateStructureContract includes a universal-axis section per rune, derived from the registry rather than hand-listedELEVATION_VALUES, PROMINENCE_VALUES, TINT_TOKENS, DENSITY_VALUES, READING_REGISTERS) rather than being restatedprominence needs a page-section header; frame/substrate need a target) are described as conditional, not as universally availablecontracts/structures.json is regenerated; the diff is additive only — no existing entry changescontract-engine-agreement.test.ts is extended to cover the new claimsrefrakt inspect --audit are checked for interactions, since both reason about which selectors existnpm run build, the full repo suite, and refrakt contracts --check passNow cheap, because each facet already owns and exports its vocabulary — that was a side effect of the SPEC-124 migration, not a goal of it.
The one genuine design question is conditionality. data-elevation can appear on any rune, but .rf-card--tinted only when a tint resolves, and data-prominence only on a rune with a page-section header. A flat "these selectors exist" list would over-promise. Describing an axis as conditional-on-config is the honest shape, and it is what the existing elements entries already do with their condition field — follow that precedent rather than inventing a second one.
This changes the contract output, which is why it was not folded into WORK-525: that item's criteria required an unchanged structures.json. Additive-only keeps it reviewable — every existing entry should be byte-identical, with new sections alongside.
Completed: 2026-09-04
Completed: 2026-09-04
Branch: claude/transform-package-refactor-7mxi8a
packages/transform/src/facets/describe.ts — UniversalAxisFacet, plus the two contract shapes it produces: UniversalAxisContract (the registry's own description of an axis, config-independent, with {block} placeholders in selector patterns) and RuneAxisContract (only what one rune's config settles).*Axis: UniversalAxisFacet beside the Facet that emits it: tint, width, content-measure, spacing, inset, density, reading, dropcap, elevation, prominence, motion, content-place, cover, frame, substrate, bg. Vocabularies are the facets' existing exports — ELEVATION_VALUES, PROMINENCE_VALUES, TINT_TOKENS, DENSITY_VALUES, READING_REGISTERS — so the contract restates nothing.facets/index.ts — UNIVERSAL_AXIS_FACETS, ordered to match ORDERED_FACETS. Also corrected a stale registry comment left by WORK-523 that still claimed tint/density/width/reading/motion/bg "resolve inline".packages/transform/src/contracts.ts — a top-level universalAxes (16 entries) and a per-rune universalAxes with axes / unavailable.contracts/structures.json + packages/lumina/contracts/structures.json regenerated: 86 KB → 185 KB, 4,224 → 7,924 lines.site/content/docs/cli/theme-tools.md — the doc that made the over-promise this item is premised on now documents both sections.Two mechanisms, matching the two kinds of gate:
prominence needs a header; tint's class needs resolved colour tokens; bg's layer needs something to draw) → a condition string on the top-level entry, following the elements[].condition precedent rather than inventing a second shape.unavailable, with the reason.There are 686 unavailability entries across the catalog — 111 runes can never carry data-reading, 111 never data-dropcap, 129 never content-place, 124 never cover, 119 never frame, 92 never data-prominence. That volume is why unavailable is a flat axis → reason map rather than a map of objects.
Emitting the full axis description on all 132 runes would have roughly tripled the file with content identical except for a block name. The axis is stated once at the top level; each rune records only its deviations — defaults, targets, and the concrete block-substituted selectors (.rf-card--tinted, .rf-card--has-bg) the original complaint named. An axis absent from both a rune's axes and its unavailable behaves exactly as the top-level entry describes.
Not by eye: strip universalAxes from the regenerated file and the entire runes subtree is byte-identical to the previous commit, key order included, as are $schema, description and prefix. The only - lines in the JSON diff are ] → ], where a key was appended after an existing one.
contract-engine-agreement.test.ts: 13 → 47 tests. Vocabularies round-tripped through the engine, {block}/{value} patterns substituted and matched, axes asserted to land on their declared target rather than the root, and every recorded unavailable driven through the engine with its inputs set.data-elevation leaves refrakt contracts --check reporting OK: up to date (132 runes) while these tests fail.facets/registry.test.ts: 3 tests pinning UNIVERSAL_AXIS_FACETS against the facet registry — coverage, no strays, and matching order. Adding a facet without describing it now fails instead of silently reopening the blind spot.refrakt inspect --audit checked for interaction: neither reads structures.json (both derive from config), so both are unaffected. Verified green regardless.- lines in the test diff are two import statements and a doc comment, all extended rather than changed.npm run build, contracts --check on both copies, full suite green at 4055 (was 4018).The generic "every recorded unavailability is honest" test first failed on cover/Chip: data-scrim present despite cover mode being unavailable. Not a bug — scrim is an input to both cover (which marks the rune root) and bg (which builds a scrim element carrying data-scrim), so a tree-wide attribute scan cannot attribute it to one facet. The check is now scoped to the root for axes that declare a target; the child-surface half is covered by the targeted reading/dropcap/frame tests. Recorded in a comment rather than worked around silently.
The motion axis declares no values. The reveal vocabulary (none|fade|slide|scale|blur) lives in the schema layer's matches, not in the engine, which passes any value through; restating it here would create a second source of truth the contract cannot keep honest. Same for width/spacing/inset, where the engine genuinely enforces no closed set — the absence of values there is accurate, not an omission.