Carry author intent without flattening it or escalating it

[chubes4]

Five fixes landed while importing one live Tailwind v4 site. Four were the same sentence:

the engine resolves author intent at a single reference viewport, then re-emits it at a cascade origin the author cannot reach.

Two independent axes that compound:

Viewport flattening — resolving at one width collapses a set of conditional values into a scalar (#1865, #1879).

Cascade escalation — the scalar is re-emitted unlayered, at raised specificity, or through Global Styles, so the author's layered rule cannot win it back (#1854, #1865, #1879).

Either alone is recoverable. Together they are unbeatable, which is why each needed a from-scratch investigation rather than pattern-matching to the last one.

Pillars

• Run parity fixtures at more than one viewport #1893 — run parity at more than one viewport (start here: makes the class visible)
• Make restate-vs-override an explicit choice in engine support CSS #1894 — restate vs override as an explicit choice (cheapest, widest blast radius)
• Give carried presentation a condition-aware representation #1895 — condition-aware carried presentation
• Treat theme.json projection as its own cascade origin #1896 — theme.json projection as its own cascade origin
• Make engine-introduced elements layout-transparent by default #1897 — engine-introduced elements layout-transparent by default

Worked examples

PR	Flattening	Escalation
#1854		layer registration order inverted
#1865	responsive font-size	layered deferral into Global Styles
#1872		(core save shape, not this pattern)
#1879	responsive display	unlayered 0,2,1 over layered author rule
#1883		(introduced box, not this pattern)

Each PR is a worked example, so the refactor can be checked against real before/after evidence rather than argued in the abstract.

---

AI assistance: analysis by Claude Sonnet 4.6 via opencode, derived from five fixes landed while importing a live Tailwind v4 site. Reviewed before filing.