Measures are a named, typed vocabulary — not a new channel

Decision

1. A registry and a prefix rule — no fourth channel

Measures are a registry of standard names with declared types, emitted through the existing config.styles. FacetResult gains nothing.

The opt-in is the prefix:

styles: { columns: '--rf-columns' }   // registered measure — validated
styles: { gap: '--bento-gap' }        // rune-local — unchecked

Any --rf--prefixed property in a styles entry must resolve to a registered measure; an unregistered one is a build error. Everything else is untouched. Zero new config surface, and the migration is "move shared concepts onto --rf-* names, leave genuinely local ones alone."

A second channel would be strictly worse: styles already carries ordered [prop, value] pairs, so a parallel measures channel doing the same thing forces the engine to arbitrate precedence between them.

2. Measures declare a type; @property registration is deferred

The registry records each measure's CSS type — --rf-index is <integer>, --rf-progress is <percentage>. That is what makes the calc() question answerable and gives the validator something to check.

Registering them with @property is a separate change. It is the right end state, but it alters runtime behaviour: registered properties gain initial values and become animatable, so an element that previously inherited nothing computes a value, and a calc() that was invalid-at-computed-value-time may start resolving. Across 132 runes that is a real rendering risk and belongs in its own change with the gallery harness covering it.

Types are data. @property is behaviour. They ship separately.

3. No positional measures

Per-child --rf-index / --rf-count, computed by the engine from a node's place among its siblings, are not built.

Their only consumer was a radial cycle, and ADR-030 concluded radial is a ladder direction rather than a topology. Rule 5b asks for a concrete design that wants it; a wheel in a plugin that does not exist is not one.

The decisive reason is different. CSS has sibling-index() and sibling-count() — the native answer to exactly this problem. Their support is narrower than :has() and wants verifying before anything relies on them, but building a per-child positional computation into the transform without first checking whether the platform already does it is the kind of thing that is regretted later. When a real consumer appears, evaluate the native functions first.

4. The admission test: does CSS compute with it?

A measure is a quantity a stylesheet computes with — it appears inside calc(), or drives a length, angle, count or proportion. If CSS only substitutes or compares the value, it is not a measure.

Applied to the current inventory:

--split-valign is the most-used custom property in the codebase (8 entries) and the test excludes it cleanly. That is not a defect finding — keyword substitution is a legitimate use of config.styles. It is evidence the boundary is real rather than drawn to fit.

5. The initial registry is small

Entries are admitted on ADR-030 rule 5a's bar — three or more existing implementations. On today's inventory that is essentially columns (five consumers, tabled above). gap has two (--bento-gap, --grid-gap) and waits. Everything else has one.

A registry of one entry looks absurd and is the correct starting point: it is the concept whose fragmentation is already costing something, and admitting more would be inventing a vocabulary rather than recording one.