Finding: SVG paint in the shared cascade

Genre: finding — grounded evidence for an open owner decision. Not a spec and not a plan. It records what was established while building the Web-first prototype, so the decision it gates can be taken on evidence rather than assumption.

Status: open as D-L in the charter's registry. No option below is chosen here.

The crux

The Web-first direction requires HTML and SVG to share one browser-grade cascade, so that a rule like .mark { fill: … } authored anywhere in the document reaches an SVG descendant through the same cascade that styles HTML. The prototype proves the cascade crosses the boundary — an HTML <style> rule reaches an inline-SVG element — but only for properties the cascade actually models. SVG paint is not among them, so today the SVG semantic compiler reads only direct fill presentation attributes outside the cascade; an inline style declaration for fill is dropped with the unknown longhand. Closing that gap is a real cost with no free option; this finding lays the options out.

Evidence

• E1 — the entire SVG paint/geometry property set is absent under the compiled engine. Stylo splits its property database by engine (servo vs gecko). The workspace compiles the servo engine; property declarations marked gecko-only are never registered, so they do not exist in the computed style. Of the 46 SVG-struct longhands in Stylo 0.16, 44 are gecko-only — the whole of fill, fill-opacity, fill-rule, stroke, stroke-width, stroke-dasharray, stroke-linecap/-linejoin/-miterlimit/-opacity/ -dashoffset, paint-order, marker-start/-mid/-end, text-anchor, clip-rule, color-interpolation(-filters), shape-rendering, vector-effect, stop-color/-opacity, flood-/lighting-color, the SVG geometry presentation properties (x,y,cx,cy,r,rx,ry,d), and the SVG mask longhands. Only clip-path and mask-image — shared box properties, not SVG paint — survive under servo. A fill declaration in a stylesheet or inline style is therefore an unknown declaration, dropped at parse time; there is no computed value to read.

• E2 — CSS custom properties do cascade and read back under servo, but as untyped token strings. A custom property set on an HTML <style> rule (.mark { --x: #16a34a }) cascades to an inline-SVG descendant with full inheritance and specificity and is readable from the computed custom-property map — verified empirically. But the value returns as the raw token string ("#16a34a"): no computed-value processing, no currentColor resolution against color, no url(#…) paint-server binding, no type checking, and custom properties inherit by default with their own invalidation semantics. The carrier reproduces cascade mechanics for a paint value; it does not reproduce SVG paint computed values.

• E3 — the gecko engine is not a viable build here. Enabling Stylo's gecko engine (which registers all the SVG longhands) pulls bindgen, mozbuild, and nsstring — the Mozilla/Gecko build system and its C++ bindings. That is not buildable in this pure-Rust Cargo workspace.

• E4 — the SVG value types themselves compile under servo. The generic, specified, and computed svg value modules (e.g. the generic SVG-paint type) are compiled unconditionally, not gecko-gated. Only the property declarations are engine-gated. A fork that un-gated the SVG paint longhands for servo is therefore plausible — but unverified at build depth, and it means carrying a patch to a large, fast-moving dependency.

• E5 — property availability is necessary but not sufficient. The proving shell currently has three independent intake gaps: its standalone entry uses HTML foreign-content parsing rather than the SVG/XML grammar; author-style collection recognizes HTML-namespace style elements but not an SVG style element; and presentation hints are not synthesized into the cascade. Un-gating fill alone would therefore still miss standalone XML, SVG-authored stylesheets, and presentation-attribute precedence.

• E6 — the current direct-attribute fallback is narrower than previously stated. It reads a fill attribute and resolves currentColor through the computed color. It does not parse fill from an inline style, so only the direct-attribute case is in hand. This distinction matters because the status quo does not cover both authoring forms.

The options

A refinement of option 3: registered custom properties (@property with a syntax, e.g. <color>) could recover typed computed values for the simple color case — but not paint-server (url(#…)) or context-dependent paint semantics, and servo @property support here is itself unverified. It narrows the loss, it does not remove it, and it needs its own spike.

Recommendation (for the owner to decide)

No option is free, so this is a genuine registry decision. The evidence suggests:

• Keep option 4 only as the proving-shell posture. It is correct for the direct attributes the prototype's fixtures actually use and is honest about what it does not do. It is not an entry into SVG-vector capability work.
• Do not adopt option 3 as the cascade of record. A carrier that reproduces cascade mechanics while discarding SVG paint semantics is, in spirit, another temporary shim — the same thing the amendment rules out when it forbids promoting the temporary SVG-only matcher. It may have a place as a narrow bridge for the cascaded-rule case, but only behind an explicit decision, never as the default.
• Scope option 2 as the path most faithful to "one browser-grade cascade." It is the only option that yields real SVG paint computed values without Gecko. Before committing to it, a timeboxed fork-feasibility spike should confirm the servo build actually computes the un-gated longhands — E4 makes it plausible, not proven.

The decision to file

D-L is registered in the charter's decision registry: how SVG paint enters the shared cascade — Gecko vs a servo-capable Stylo fork vs a scoped custom-property carrier vs direct-attribute status quo. Its evidence bar is this finding plus a bounded feasibility bundle that proves the full ingress, not merely a longhand build: the SVG/XML grammar entry, SVG-namespace stylesheet intake, presentation-hint precedence, minimal paint-longhand computation, and precedence/inheritance/currentColor/invalid-value behavior. Until the owner decides it, the Web-first path reads direct paint attributes outside the cascade, says so, and does not accrete SVG-vector capability on that scaffold.