An edit to the runtime's 65k lines spent 35-90 s in opt-level 1 codegen before
its test binary linked. The hot pipeline is generic and compiles in the
dependants anyway; theme, behavior and proof stay at opt-level 1.
cargo unifies features only over the packages a command selects, so a framework
crate's tests, a domain crate's tests, the gallery and the story renderer each
built their own copy of the framework and of wgpu. feature-root carries the
gallery's dependencies with every domain crate; cl selects it beside every
command's package and names the crate's own targets, so nothing of it builds.
The per-crate feature anchors go; cl check also checks a crate's bins.
cl test failed for craft-ui-panels, canvas, patterns and shell: their canonical set named craft-ui-contract/snapshot, and cargo lets a command name a feature only of a direct dependency, which the runtime is not for crates above widgets. craft-ui-contract's testing now includes snapshot (the resolved feature set of the runtime in a test build is the same three features as before, so no unit changes), and the canonical sets drop the snapshot entries. cl doctor resolves every crate's canonical set against cargo metadata, with no build: a plain feature must be the crate's own, a dep/feature must name a direct dependency and a feature it has. docs/development.md states what CI covers (ui.yml for the seven ui crates, theme.yml, docs.yml, path-filtered workflows for some leaf crates) and what it does not (the domain crates on the framework, the gallery, the renderer, the umbrella crate, any workspace-wide build).