makeover_timing/

lib.rs

//! The time axis of the make-family design system.
//!
//! <!-- wiki: makeover-timing -->
//!
//! [`makeover`] resolves colour, `makeover-geometry` resolves distance, and
//! `makeover-layout` names what a thing is without resolving anything. This
//! crate answers the remaining question of the same shape: **how long**.
//!
//! The move is the one the family makes everywhere. A duration is named by
//! what it is waiting for, and the number follows:
//!
//! | Intent | What it waits for | Resolves to |
//! |---|---|---|
//! | [`Intent::Revert`] | a control returning to its resting label after confirming | 1500ms |
//! | [`Intent::Clear`] | a status line emptying itself | 2000ms |
//! | [`Intent::Dismiss`] | a transient notice's lifetime | 3000ms |
//! | [`Intent::Debounce`] | typing settling before the work starts | 150ms |
//!
//! Whether a toast should live 3000ms or 3500ms is unanswerable on its own.
//! Whether a message is a toast or a banner is not. That is the whole argument
//! for the layer, and it is the same one [`Gap`](makeover_geometry::Gap) makes
//! about six pixels.
//!
//! # One duration per intent, on every renderer
//!
//! An intent resolves to exactly one duration everywhere. Not one per renderer,
//! not one per theme.
//!
//! A per-renderer table, resolving the way
//! [`Density`](makeover_geometry::Density) resolves gaps, is the wrong shape.
//! Distance has a renderer axis because a fingertip is coarser than a cursor
//! and a terminal cell is coarser than a pixel: the *surface* differs. Time
//! does not differ that way. A second is a second in a browser, in egui and in
//! a terminal, and the reader waiting it out is the same reader.
//!
//! So a divergence here is a bug report, not an axis.
//!
//! # What is deliberately not a timing intent
//!
//! Three classes are out of scope, and leaving them out is most of what makes
//! the four above coherent.
//!
//! **A race is not an intent.** Waiting 300ms before navigating because the
//! write "should" have landed, or 150ms before closing a dropdown so a
//! mousedown can beat the blur, is a synchronisation bug wearing a duration's
//! clothes. Naming those would launder them into design decisions and give
//! every future one a token to hide behind. They get fixed per screen.
//!
//! **Severity is not an intent either, and it is not a fifth number.** Two
//! renderers already reached for one: MNW gives an error toast 6000ms against
//! an ordinary one's 3000ms, and audiofiles' footer never expires an error at
//! all while an ordinary message goes at 30s. Read together those are not two
//! durations, they are one statement — *a message the user must not miss does
//! not go away on its own* — and `makeover-layout` already has the word for
//! it: such a message is not `Notice::transient`. It is a banner, and a banner
//! has no lifetime. See [`notice_lifetime`].
//!
//! **A poll interval is not an intent.** A retry backoff, a health check, an
//! update check: those are answerable from what they talk to, not from what a
//! reader can follow, and nothing here has an opinion about them.
//!
//! # Motion is a separate axis
//!
//! [`Intent`] says how long a state lasts. [`Motion`] says how long a change
//! takes. CSS itself draws that line — a `setTimeout` against a
//! `transition-duration` — and folding the two together is what makes a
//! "timing scale" unusable: 300ms of fade and 3000ms of toast are not two
//! rungs of one ramp, they are answers to different questions.
//!
//! [`Motion`] has one rung today because the tree has one measured transition.
//! It is an enum rather than a constant so the second one has somewhere to go,
//! and it grows when something is measured, not when a scale looks short.
//!
//! # Cadence is a third question
//!
//! [`Intent`] is how long a state lasts and [`Motion`] is how long a change
//! takes. Both are one-shot: something starts, it ends. [`Cadence`] is how
//! often a repeating mark repeats, which neither of the other two can answer
//! without lying about its own shape.
//!
//! The crate header's exclusion of poll intervals does not cover it, and the
//! difference is the same one that exclusion rests on. A backoff or a health
//! check is answerable from what it talks to. How fast a mark may blink before
//! it reads as an alarm is answerable from the reader, which is what every rung
//! in this crate is answerable from.
//!
//! # Reduced motion
//!
//! The first thing in the family to have a motion-off path, opened here
//! because [`Cadence`] is the first token whose whole existence is movement.
//! A duration that is not animating anything is unaffected by it:
//! [`Intent::Debounce`] is a wait, not a stroke, and reducing motion does not
//! make input settle faster.
//!
//! Two halves, because a browser and an egui window learn about the preference
//! differently. A web surface gets a `prefers-reduced-motion` block in the
//! generated stylesheet and needs no code. Every other renderer asks
//! [`activity_blink`] with the bool its own platform gave it, exactly as it
//! asks [`notice_lifetime`] with the bool the description gave it.
//!
//! # Where the numbers came from
//!
//! Every value below is a count from the tree, not a preference:
//!
//! ```text
//! revert     1500ms   MNW: 6 hand-rolled sites, plus core/clipboard.ts's own default
//! clear      2000ms   MNW: 4 sites
//! dismiss    3000ms   MNW: the toast renderer's lifetime
//! debounce    150ms   audiofiles SEARCH_DEBOUNCE, MNW docs-search.js
//! fade        300ms   MNW: TOAST_FADE_MS, matching the .fade-out transition
//! ```
//!
//! The one contested value is the debounce, where MNW's two category
//! typeaheads sit at 200ms against everything else's 150ms. 150 wins on the
//! count and on the cross-renderer agreement, and the 200s conform.
//!
//! [`Cadence::Activity`] is the one rung not counted off the tree. Wiki
//! `loading-and-progress-standard` rules that an unmeasured wait blinks rather
//! than spins; this is the cadence it blinks at. What it is answerable to is
//! stated on the member.
//!
//! # Consumers
//!
//! Web surfaces bake [`timing_css`] in at build time. Nothing here changes at
//! runtime, so there is no load-time JS step, exactly as with geometry. egui
//! and ratatui surfaces read [`Intent::duration`] instead, which is why this
//! is a crate rather than a stylesheet.
//!
//! [`makeover`]: https://makenot.work/git/max/makeover

#![forbid(unsafe_code)]

use std::fmt::Write as _;
use std::time::Duration;

use makeover_geometry::in_css_layer;

/// A duration named by what it is waiting for.
///
/// Four members, and the set is closed on purpose: each one is a thing a
/// reader is waiting through, and the crate header says what was measured out.
/// Adding a fifth means naming a wait nobody here is already having.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, PartialOrd, Ord)]
pub enum Intent {
    /// How long a control shows that it did something before returning to its
    /// resting label.
    ///
    /// "Copied!" on a button that said "Copy link". Long enough to be read
    /// after the eye has moved back to it, short enough that the control is
    /// honest about its own label again before the next click.
    Revert,
    /// How long a status line holds a message before emptying itself.
    ///
    /// The message is a receipt for something the user just did, so it is read
    /// or not read immediately. Holding it longer means the next glance at that
    /// line reports stale news.
    Clear,
    /// How long a transient notice lives before it starts to leave.
    ///
    /// Excludes the leaving itself, which is [`Motion::Fade`]. A notice the
    /// user must not miss is not transient and gets no lifetime at all; see
    /// [`notice_lifetime`].
    Dismiss,
    /// How long input waits to settle before the work behind it starts.
    ///
    /// A search field that queries on every keystroke, filtered through this.
    /// The number is a claim about typing rather than about the query: below
    /// roughly 100ms an ordinary typist trips it mid-word, and above roughly
    /// 250ms the field feels like it stopped listening.
    Debounce,
}

impl Intent {
    /// The duration in whole milliseconds.
    ///
    /// The primary resolution. [`Self::duration`] and [`Self::css`] are both
    /// spellings of this number, so there is exactly one place it lives.
    #[must_use]
    pub const fn ms(self) -> u32 {
        match self {
            Self::Revert => 1500,
            Self::Clear => 2000,
            Self::Dismiss => 3000,
            Self::Debounce => 150,
        }
    }

    /// The duration as a [`Duration`], for the renderers that are not a
    /// browser.
    #[must_use]
    pub const fn duration(self) -> Duration {
        Duration::from_millis(self.ms() as u64)
    }

    /// The CSS custom property name, without the leading dashes.
    #[must_use]
    pub const fn token(self) -> &'static str {
        match self {
            Self::Revert => "timing-revert",
            Self::Clear => "timing-clear",
            Self::Dismiss => "timing-dismiss",
            Self::Debounce => "timing-debounce",
        }
    }

    /// The CSS value, as a `ms` time.
    ///
    /// Milliseconds rather than seconds at every rung, including the ones that
    /// divide evenly: a stylesheet where some durations read `1.5s` and others
    /// `150ms` cannot be scanned for the odd one out.
    #[must_use]
    pub fn css(self) -> String {
        format!("{}ms", self.ms())
    }

    /// Every intent, in the order they are emitted.
    #[must_use]
    pub const fn all() -> [Self; 4] {
        [Self::Revert, Self::Clear, Self::Dismiss, Self::Debounce]
    }
}

/// How long a change takes, as opposed to how long a state lasts.
///
/// See the crate header for why this is not a fifth [`Intent`]. One rung, and
/// it grows from a measurement rather than from the scale looking short.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, PartialOrd, Ord)]
pub enum Motion {
    /// A thing leaving: opacity to zero, then gone.
    ///
    /// The whole of the departure. A renderer that removes the node itself has
    /// to wait this out first, or it drops the animation mid-stroke.
    Fade,
}

impl Motion {
    /// The duration in whole milliseconds.
    #[must_use]
    pub const fn ms(self) -> u32 {
        match self {
            Self::Fade => 300,
        }
    }

    /// The duration as a [`Duration`].
    #[must_use]
    pub const fn duration(self) -> Duration {
        Duration::from_millis(self.ms() as u64)
    }

    /// The CSS custom property name, without the leading dashes.
    #[must_use]
    pub const fn token(self) -> &'static str {
        match self {
            Self::Fade => "motion-fade",
        }
    }

    /// The CSS value, as a `ms` time.
    #[must_use]
    pub fn css(self) -> String {
        format!("{}ms", self.ms())
    }

    /// Every motion, in the order they are emitted.
    #[must_use]
    pub const fn all() -> [Self; 1] {
        [Self::Fade]
    }
}

/// How often a repeating mark repeats.
///
/// The third question, and see the crate header for why it is neither an
/// [`Intent`] nor a [`Motion`]. One rung, and it grows the same way the others
/// do.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, PartialOrd, Ord)]
pub enum Cadence {
    /// How long the activity mark holds each of its two states.
    ///
    /// The hard-disk light: a small mark saying that something is happening,
    /// on a wait with no countable size. `makeover-layout`'s `Awaiting` is what
    /// says the wait has no size; this is how fast the answer to that blinks.
    ///
    /// **A half-period, not a cycle.** The mark is lit for this long, dark for
    /// this long, and a full cycle is twice it. One number rather than a period
    /// plus a duty cycle, because two numbers are two things three renderers
    /// can disagree about, and a mark that is lit a fifth of the time in a
    /// browser and half the time in a terminal is not one mark.
    ///
    /// 500ms, so a cycle is a second. Not a count off the tree like every other
    /// rung, since nothing has drawn this yet. What it is answerable to is the
    /// eye: fast enough that a glance catches it working, slow enough that it
    /// reads as steady work rather than as an alarm. A blink much under half a
    /// second is a strobe and starts to claim urgency the wait has not earned.
    ///
    /// This is the *fallback*, and on most surfaces it should be the rarer
    /// case. Where the activity is observable the mark follows it, per rule 3
    /// of wiki `loading-and-progress-standard`: an upload blinks per chunk
    /// delivered, and this is for the waits with nothing to watch.
    Activity,
}

impl Cadence {
    /// The duration in whole milliseconds.
    #[must_use]
    pub const fn ms(self) -> u32 {
        match self {
            Self::Activity => 500,
        }
    }

    /// The duration as a [`Duration`].
    #[must_use]
    pub const fn duration(self) -> Duration {
        Duration::from_millis(self.ms() as u64)
    }

    /// One full on-and-off cycle, which is twice the half-period.
    ///
    /// Spelled here rather than doubled at each call site: a renderer driving a
    /// two-state toggle wants [`duration`](Self::duration), and one scheduling
    /// a whole cycle wants this, and neither should be doing the arithmetic.
    #[must_use]
    pub const fn cycle(self) -> Duration {
        Duration::from_millis(self.ms() as u64 * 2)
    }

    /// The CSS custom property name, without the leading dashes.
    #[must_use]
    pub const fn token(self) -> &'static str {
        match self {
            Self::Activity => "cadence-activity",
        }
    }

    /// The CSS value, as a `ms` time.
    #[must_use]
    pub fn css(self) -> String {
        format!("{}ms", self.ms())
    }

    /// Every cadence, in the order they are emitted.
    #[must_use]
    pub const fn all() -> [Self; 1] {
        [Self::Activity]
    }
}

/// How fast the activity mark blinks, given whether the reader has asked for
/// less motion.
///
/// The same seam as [`notice_lifetime`], and it takes a bool for the same
/// reason: the preference is the platform's to report, and taking the answer
/// rather than the platform keeps this crate off everyone's dependency graph.
///
/// ```
/// # use makeover_timing::{Cadence, activity_blink};
/// assert_eq!(activity_blink(false), Some(Cadence::Activity.duration()));
/// assert_eq!(activity_blink(true), None);
/// ```
///
/// `None` is not "the caller decides" and it is not "draw nothing". It means
/// the mark does not blink: it is drawn, lit, and still, for as long as the
/// wait lasts. The reader still learns that something is happening, which is
/// the whole content of the mark; what they are spared is the movement. A
/// renderer that hides the mark instead has removed the information rather than
/// the animation, and reduced motion asks for the second.
///
/// A web surface does not call this. The generated stylesheet carries a
/// `prefers-reduced-motion` block that does the same thing in the cascade,
/// which is why [`reduced_motion_css`] exists.
#[must_use]
pub const fn activity_blink(reduced: bool) -> Option<Duration> {
    if reduced {
        None
    } else {
        Some(Cadence::Activity.duration())
    }
}

/// How long a notice lives, given whether the description calls it transient.
///
/// The seam this crate was built for. `makeover-layout` documents
/// `Notice::Toast` as "transient, stacked, dismisses itself" and
/// `Notice::transient()` returns true for it — the description says a notice
/// goes away on its own and deliberately says nothing about when. This is
/// when, on the renderer's side of the line:
///
/// ```
/// # use makeover_timing::{Intent, notice_lifetime};
/// // notice_lifetime(makeover_layout::Notice::Toast.transient())
/// assert_eq!(notice_lifetime(true), Some(Intent::Dismiss.duration()));
/// assert_eq!(notice_lifetime(false), None);
/// ```
///
/// Taking the bool rather than the enum is what keeps this crate off
/// `makeover-layout`'s dependency graph, and the bool is exactly what the
/// description asserts. A renderer already holds the `Notice`.
///
/// `None` is not "the caller decides". It means the notice has no lifetime:
/// a banner is dismissed by fixing the cause, and an error is a banner. See
/// the crate header on why severity is not a fifth duration.
#[must_use]
pub const fn notice_lifetime(transient: bool) -> Option<Duration> {
    if transient {
        Some(Intent::Dismiss.duration())
    } else {
        None
    }
}

/// Emit the time axis as CSS declarations, no selector.
///
/// [`Intent::Debounce`] is emitted with the rest even though no stylesheet can
/// use it. The point of the layer is that one document holds every duration in
/// the system; a token that lives here for three of the four intents and in a
/// JS constant for the fourth is the drift this crate exists to end, and a
/// script can read the value back off the computed style.
#[must_use]
pub fn timing_css_declarations() -> String {
    let mut out = String::new();
    out.push_str("  /* Time. Named for what is being waited on; the number\n");
    out.push_str("     follows. One duration per intent on every renderer —\n");
    out.push_str("     a divergence here is a bug report, not an axis. */\n");
    for intent in Intent::all() {
        let _ = writeln!(out, "  --{}: {};", intent.token(), intent.css());
    }
    out.push_str("\n  /* Motion: how long a change takes, not how long a state\n");
    out.push_str("     lasts. A separate question, so a separate axis. */\n");
    for motion in Motion::all() {
        let _ = writeln!(out, "  --{}: {};", motion.token(), motion.css());
    }

    out.push_str("\n  /* Cadence: how often a repeating mark repeats. A\n");
    out.push_str("     half-period, so a full cycle is twice it. */\n");
    for cadence in Cadence::all() {
        let _ = writeln!(out, "  --{}: {};", cadence.token(), cadence.css());
    }
    out
}

/// The motion-off block, for a reader who has asked for less of it.
///
/// Everything here that animates resolves to `0ms`. A zero-length transition
/// is a jump to its end state, so a notice stops fading and simply goes when
/// its lifetime is up, with no rule written twice.
///
/// **The blink needs one thing from the rule that consumes it**, and it does
/// not fall out of the zero on its own. A zero-length animation with no
/// `animation-fill-mode` leaves the element in its *base* style, not at its
/// last keyframe. So write the rule with the mark lit in the base style and the
/// keyframes doing the dimming, never the other way round. Then reduced motion
/// stills a lit mark, and the inverted spelling would blank it. See
/// [`activity_blink`] on why removing the mark answers a different request from
/// the one that was made.
///
/// [`Intent`] is untouched. Those are waits rather than strokes, and a reader
/// asking for less motion has not asked for their input to settle sooner or for
/// a notice they are reading to leave early.
#[must_use]
pub fn reduced_motion_css() -> String {
    let mut out = String::new();
    out.push_str("@media (prefers-reduced-motion: reduce) {\n");
    out.push_str("  :root {\n");
    out.push_str("    /* Motion off. The state a stroke was heading for,\n");
    out.push_str("       reached at once; nothing is removed. */\n");
    for motion in Motion::all() {
        let _ = writeln!(out, "    --{}: 0ms;", motion.token());
    }
    for cadence in Cadence::all() {
        let _ = writeln!(out, "    --{}: 0ms;", cadence.token());
    }
    out.push_str("  }\n}\n");
    out
}

/// Emit the whole time axis: a `:root { … }` block, then the motion-off block.
///
/// Mirrors `makeover_geometry::geometry_css_vars`. Like geometry and unlike
/// colour, none of this varies at runtime, so a web consumer bakes it in at
/// build time rather than applying it from JS on load.
///
/// The `prefers-reduced-motion` block rides with the values it overrides rather
/// than being a second thing to remember to include. A consumer that took the
/// vars and not the block would animate at every rung for a reader who asked it
/// not to, and would do it silently.
#[must_use]
pub fn timing_css_vars() -> String {
    format!(
        ":root {{\n{}}}\n\n{}",
        timing_css_declarations(),
        reduced_motion_css()
    )
}

/// The time axis as a stylesheet, inside the family's cascade layer.
///
/// The whole-file entry point, and the one a build script should call.
/// Unlayered declarations outrank every named layer, so generated CSS that
/// stays outside the layer beats the app's own overrides regardless of
/// specificity — which is invisible until the app adopts layers, and then is a
/// puzzle. `makeover_geometry::CSS_LAYER` is the one spelling of the name.
#[must_use]
pub fn timing_css() -> String {
    in_css_layer(&timing_css_vars())
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn every_intent_resolves_to_one_number_in_three_spellings() {
        // ms, Duration and CSS are three renderings of one value, so a rung
        // cannot drift between the browser and egui.
        for intent in Intent::all() {
            assert_eq!(intent.duration().as_millis() as u32, intent.ms());
            assert_eq!(intent.css(), format!("{}ms", intent.ms()));
        }
        for motion in Motion::all() {
            assert_eq!(motion.duration().as_millis() as u32, motion.ms());
        }
        for cadence in Cadence::all() {
            assert_eq!(cadence.duration().as_millis() as u32, cadence.ms());
            assert_eq!(cadence.css(), format!("{}ms", cadence.ms()));
        }
    }

    #[test]
    fn the_measured_values_are_the_ones_the_tree_had() {
        // Pinned against the 2026-08-18 count. Changing one of these is a
        // design decision about every consumer at once, which is the point of
        // the crate; a test failure is the argument happening out loud.
        assert_eq!(Intent::Revert.ms(), 1500);
        assert_eq!(Intent::Clear.ms(), 2000);
        assert_eq!(Intent::Dismiss.ms(), 3000);
        assert_eq!(Intent::Debounce.ms(), 150);
        assert_eq!(Motion::Fade.ms(), 300);
        // The one value that is not a count. Pinned all the same: it is the
        // number three renderers agree on, which is the whole reason it is
        // here rather than in each of them.
        assert_eq!(Cadence::Activity.ms(), 500);
    }

    #[test]
    fn a_notice_leaves_after_its_lifetime_and_its_fade() {
        // The two numbers the toast class needs, and the reason they are on
        // different axes: a renderer that removes the node at Dismiss drops
        // the animation, and one that waits Dismiss + Fade is correct.
        assert_eq!(notice_lifetime(true), Some(Duration::from_secs(3)));
        assert!(Motion::Fade.duration() < Intent::Dismiss.duration());
    }

    #[test]
    fn a_notice_that_is_not_transient_has_no_lifetime() {
        // Not "the caller decides" — a banner is dismissed by fixing the cause.
        // This is where an error toast's second number went.
        assert_eq!(notice_lifetime(false), None);
    }

    #[test]
    fn debounce_is_the_shortest_wait_and_a_notice_the_longest() {
        // The ordering is the sanity check on the set: input settling is the
        // one wait a user is inside rather than watching, so it is the only
        // sub-second rung, and nothing may quietly grow past a notice.
        assert!(
            Intent::all()
                .iter()
                .all(|i| i.ms() >= Intent::Debounce.ms())
        );
        assert!(Intent::all().iter().all(|i| i.ms() <= Intent::Dismiss.ms()));
    }

    #[test]
    fn no_intent_is_long_enough_to_be_a_poll_interval() {
        // A ceiling with an argument behind it: every rung here is a wait a
        // reader sits through, and past a few seconds that stops being true.
        // A backoff or a health check answers to what it talks to, not here.
        assert!(Intent::all().iter().all(|i| i.ms() <= 5_000));
    }

    #[test]
    fn the_layer_is_emitted_inside_the_family_layer() {
        let css = timing_css();
        assert!(css.starts_with("@layer makeover {\n"));
        assert!(css.contains("    :root {"));
        assert!(css.trim_end().ends_with('}'));
    }

    #[test]
    fn every_token_reaches_the_stylesheet_exactly_once() {
        let css = timing_css();
        for intent in Intent::all() {
            let decl = format!("--{}: {}", intent.token(), intent.css());
            assert_eq!(css.matches(&decl).count(), 1, "{}", intent.token());
        }
        for motion in Motion::all() {
            let decl = format!("--{}: {}", motion.token(), motion.css());
            assert_eq!(css.matches(&decl).count(), 1, "{}", motion.token());
        }
        for cadence in Cadence::all() {
            let decl = format!("--{}: {}", cadence.token(), cadence.css());
            assert_eq!(css.matches(&decl).count(), 1, "{}", cadence.token());
        }
    }

    #[test]
    fn tokens_are_prefixed_by_their_axis() {
        // `--timing-*` for a state's length, `--motion-*` for a change's. A
        // reader scanning the sheet can tell which question a var answers.
        assert!(
            Intent::all()
                .iter()
                .all(|i| i.token().starts_with("timing-"))
        );
        assert!(
            Motion::all()
                .iter()
                .all(|m| m.token().starts_with("motion-"))
        );
        assert!(
            Cadence::all()
                .iter()
                .all(|c| c.token().starts_with("cadence-"))
        );
    }

    #[test]
    fn the_blink_is_a_half_period_and_the_cycle_is_twice_it() {
        // The one arithmetic a renderer must not be doing itself. A mark lit
        // for 500ms and dark for 500ms is a one-second cycle, and a renderer
        // scheduling the cycle where it meant the half draws at half speed.
        assert_eq!(Cadence::Activity.cycle(), Cadence::Activity.duration() * 2);
        assert_eq!(Cadence::Activity.cycle(), Duration::from_secs(1));
    }

    #[test]
    fn the_blink_is_slower_than_a_stroke_and_faster_than_a_notice() {
        // Where it sits between the existing axes is the sanity check on the
        // number. Faster than Fade and it is a strobe; slower than the shortest
        // thing a reader sits through and it reads as stalled rather than busy.
        assert!(Cadence::Activity.duration() > Motion::Fade.duration());
        assert!(Cadence::Activity.duration() < Intent::Revert.duration());
    }

    #[test]
    fn reduced_motion_stills_the_mark_rather_than_removing_it() {
        // None means lit and static, not absent. The distinction is the whole
        // of what reduced motion asks for: less movement, not less information.
        assert_eq!(activity_blink(false), Some(Cadence::Activity.duration()));
        assert_eq!(activity_blink(true), None);
    }

    #[test]
    fn reduced_motion_leaves_the_waits_alone() {
        // Only what animates is zeroed. A debounce is a wait, and a reader
        // asking for less motion has not asked their input to settle sooner.
        let css = reduced_motion_css();
        for intent in Intent::all() {
            assert!(!css.contains(intent.token()), "{}", intent.token());
        }
        for motion in Motion::all() {
            assert!(css.contains(&format!("--{}: 0ms;", motion.token())));
        }
        for cadence in Cadence::all() {
            assert!(css.contains(&format!("--{}: 0ms;", cadence.token())));
        }
    }

    #[test]
    fn the_motion_off_block_ships_with_the_values_it_overrides() {
        // A consumer taking the vars and not the block animates at every rung
        // for a reader who asked it not to, and does it silently. So the
        // whole-file entry point carries both, inside the family layer.
        let css = timing_css();
        assert!(css.contains("@media (prefers-reduced-motion: reduce)"));
        let vars = css
            .find(&format!(
                "--{}: {}",
                Cadence::Activity.token(),
                Cadence::Activity.css()
            ))
            .expect("the value");
        let off = css.find("prefers-reduced-motion").expect("the block");
        assert!(
            vars < off,
            "the override has to come after what it overrides"
        );
    }

    #[test]
    fn no_two_rungs_share_a_name_or_a_value() {
        // A duplicate name silently overwrites in the cascade; a duplicate
        // value is two names for one thing, which is a distinction nobody can
        // choose between.
        let mut tokens: Vec<&str> = Intent::all().iter().map(|i| i.token()).collect();
        tokens.extend(Motion::all().iter().map(|m| m.token()));
        tokens.extend(Cadence::all().iter().map(|c| c.token()));
        let mut sorted = tokens.clone();
        sorted.sort_unstable();
        sorted.dedup();
        assert_eq!(sorted.len(), tokens.len(), "{tokens:?}");

        let mut values: Vec<u32> = Intent::all().iter().map(|i| i.ms()).collect();
        values.sort_unstable();
        values.dedup();
        assert_eq!(values.len(), Intent::all().len(), "two intents, one number");
    }
}