ADR-0007: Export Regular and Clear as built-in Glass styles

Haze will expose GlassStyle.regular and GlassStyle.clear as shared built-in styles rather than adding a public modifier variant parameter or extending the public GlassOptics API with semantic cases. The parallel sample-only preset catalogue will be removed rather than retained or exported. Regular remains the default size-responsive material. Clear initially used size-responsive blur and depth with an authored refraction response, and remains recognisably distinct when a renderer simplifies advanced optics. Both styles own the complete optical, edge, lighting, chromatic, tone, and content-normal response while preserving independently composed shape, background colour, tint, alpha, light position, and interaction appearance. These are Haze styles informed by iOS's semantic distinction and measured direction, not promises of pixel parity or Apple-internal constants. Identity remains outside this decision because disabling attachment and rendering is modifier/runtime behaviour rather than a Glass appearance.

Consequences

Callers select a built-in response through the existing replayable Style seam and customize it with then, without learning another public type or precedence rule. Switching between Regular and Clear resets every built-in-style-owned channel but does not erase theme or caller presentation. Samples use Regular and Clear directly; sample-only Adaptive, Clear, Frosted, Deep, and Prism presets are removed. A sample that teaches custom authoring declares that customization locally instead of adding another reusable preset catalogue.

ADR-0008 revises the representation of their optics: Regular and Clear remain built-in Glass styles, but their optical behavior is expressible through the same public configuration available to callers.

Diffusion calibration, 2026-09-16

Clear now uses constant shallow diffusion: OpticalSizeValue.Fixed(1f) for depth and OpticalSizeValue.Fixed(1.25.dp) for blur radius. This supersedes the size-responsive Clear response above. Native UIKit captures on iOS 27.0 (24A434), built with Xcode 27.0 (27A266a) and SDK 27.0, use the same 64/176/220pt geometry and 8pt grid as Haze's size fixture. The simulator's Liquid Glass preference is Default (Tint Amount 50%); Reduce Transparency and Increase Contrast are disabled. Clear retains approximately 23/27/28% of the unfiltered grid's adjacent-pixel linear-luminance contrast. Light and dark Clear captures agree. This replaces the initial iOS 26 calibration, whose more diffuse Clear reference led to a provisional 2dp radius.

Regular retains approximately 24% on the compact capsule and effectively none on the larger native surfaces. An independent launch reproduces that response. Haze keeps its compact diffusion and increases the 176/220dp depth anchors to 0.98/1, reducing the sharp-source contribution on cards and removing it on panels. Its 64dp depth remains 0.85 and its blur radii remain 4/10/15dp.

These are empirical diffusion targets, not Apple-specified radii or a pixel-parity contract. The metric includes tone and refraction as well as blur, and native Regular's adaptive appearance and edge treatment differ. The calibration prioritizes stronger diffusion on larger Regular surfaces. Both styles continue to use the existing public optics configuration; no renderer-specific identity branch or additional API is needed.

Clear refraction calibration, 2026-09-16

A separate four-phase native capture isolates apparent sampling displacement from the grid's repeating lines. Clear now selects RefractionProfile.Edge(28.dp), with strength 0.85 and authored displacement 56dp. The fitted inward decay and optical corner radius (1.5 times the visible radius, capped at half the shortest side) reproduce the accepted corner loops. Displacement is capped at half the shortest side. The production normal is analytic rather than four additional distance samples per pixel. Lighting retains Circle, with height fraction 0.35; white point is 0.17.

RefractionProfile.Surface remains the default for caller-authored optics. This explicit profile preserves existing surface semantics; Edge width controls refraction independently of the surface height used for lighting. Both base and interaction detail use the same sampling map. Zero width disables Edge sampling and its detail pass.

These settings are an empirical visual match for the accepted fixture, not Apple's implementation or a promise of parity across content.

Regular refraction and wide diffusion calibration, 2026-09-16

Further grid and photographic comparisons supersede the provisional Regular settings above. Regular now uses Edge(20.dp) with authored 48dp displacement at 0.7 strength, lighting height fraction 0.15, and no secondary detail or fold. Depth is 0.65/1/1 and blur is 20/24/25dp at 64/176/220dp. Compact controls retain refracted grid loops; cards and panels remove both fine grid lines and medium-scale photographic detail. The light-appearance tone uses white point 0.55 and chroma multiplier 1.5, with ambient response 0.15 and edge softness 1dp. Clear's accepted settings remain unchanged.

This exposed a source-space blur cap of 38.5 physical pixels: larger authored radii previously produced the same result at dense resolutions. Non-progressive blur now supports up to 256px. Both backends use native Gaussian filtering for uniform blur above four input pixels so input bounds include all required samples, including lower-density Performance rendering. The custom runtime-filter paths exposed sharp backdrop strips near output boundaries. Android retains one fused output graph; the retained Skiko path caches the native filter and its blurred layer. Small-radius and progressive blur keep their semantic kernels, and progressive blur retains its previous radius bound. Layer budgets and lifecycle follow the selected backend path. This wide-blur correctness exception to ADR-0003 has host pixel coverage; its physical-device performance has not been measured.

Native light and dark appearances differ substantially. The original fixed Haze response targeted the light reference; subsequent system appearance support selects separate light and dark responses at node resolution. The comparisons establish visual direction, not physical GPU performance or native content adaptation.

System appearance response, 2026-09-23

Regular and Clear remain shared, immutable Style values. A built-in records an appearance-dependent write in its ordered replay sequence, and the attached Glass node reads the Compose host's system appearance. This also applies to the implicit Regular response when style is omitted. A live host appearance change replays the sequence and invalidates affected rendering without reconstructing caller Styles. Later writes retain precedence, and a later built-in resets its owned channels while preserving shape, background, tint, alpha, light position, and interactions.

The light response retains the 2026-09-16 calibration. The dark response lowers white point and ambient lift and changes edge and highlight balance while retaining Regular's size-dependent diffusion and Clear's shallow, more transparent optics. These are Haze approximations, not iOS pixel matches or Apple coefficients. Material 3 remains a separate optional builder for app-theme surface color and tint. The host may override device appearance; an unknown Skiko theme uses the light response. The pinned Skiko theme local is internal and deprecated, so Compose upgrades must recheck that adapter.

The new dark references use a 64/176/220dp capsule/card/panel fixture at 1dp per pixel, with a structured dark grid for Regular and a bright photograph for Clear. Desktop captures use a 420×800px Skiko host; Android host captures use the Pixel 5 Robolectric profile at SDK 28, 32, and 35. Each style also passes a light→dark→light capture on the same attached node: more than 1% of pixels change, and the restored image differs from the initial image by less than one 8-bit color step on average. The Android SDK 28 path exercises fallback rendering. These checks establish Haze output and reversibility under controlled host appearance; they do not establish physical-device behavior, a manual OS switch, or a side-by-side native dark match.

Colour handling, 2026-09-17

chromaMultiplier above 1f is a requested chroma boost, bounded for source colours already in the sRGB gamut. A boost that would leave that gamut is reduced to the in-gamut limit, and an already out-of-gamut source is left unchanged. The purpose is source-colour preservation: clamping an amplified saturated colour would create a broad colour plateau that loses the original hue and detail.

Authored tint and edgeShadow retain Compose ColorSpace semantics through colour-managed shader uniforms on supported renderers. This decision does not settle the whole colour model: encoded luma() used to construct content normals, display-space contrast, and ambient multiplication remain explicitly deferred questions rather than claims of scene-linear behaviour.