makeover-touch 0.3.0

//! The adaptation layer of the make-family design system.
//!
//! <!-- wiki: makeover-touch -->
//!
//! `makeover` answers *what colour*. `makeover-geometry` answers *how much
//! space*, and owns the two axes an adaptation is stated against:
//! [`Density`] (pointer or touch) and [`SizeClass`] (compact, medium,
//! expanded). `makeover-layout` answers *what the thing is*. This crate
//! answers one question and no other:
//!
//! > Does this affordance exist here?
//!
//! Like `makeover-layout` it emits nothing. It is a description, rendered to
//! CSS by `makeover-webview` and to whatever the other renderers can express.
//!
//! # Why this is a crate and not a density preset
//!
//! Measured across the MNW server's `@media` blocks (137) and goingson's
//! `ui-mode-*` blocks (192), bucketed by what the declarations inside actually
//! change:
//!
//! | bucket | MNW | GO | retired by |
//! |---|---|---|---|
//! | density | 32% | 27% | a `makeover-geometry` preset |
//! | type | 23% | 15% | the type scale |
//! | columns | 20% | 12% | `makeover_layout::Column` |
//! | reflow | 16% | 17% | `makeover_layout::Arrangement` |
//! | **show/hide** | **12%** | **12%** | **this crate** |
//! | **reposition** | **6%** | **20%** | **this crate** |
//! | **appearance** | **1%** | **17%** | **this crate** |
//!
//! The bottom three are the roughly 43% that no spacing scale can retire, and
//! trying is the mistake this whole family already made once. `display: none`
//! on a keyboard hint says *the affordance does not exist on touch*. No amount
//! of gap retuning expresses that, and a scale that tried would be smuggling a
//! product claim onto a measurement axis — which is exactly what the 2026-07-29
//! Touch demolition was for.
//!
//! Two of the eight members are not in that census at all, and the exception is
//! worth stating rather than leaving to be noticed. [`Affordance::Gesture`] and
//! [`Affordance::Haptic`] adapt *behaviour*, which no stylesheet contains, so
//! counting `@media` blocks could never have found them. Their evidence has the
//! same shape in a different file: goingson `frontend/js/touch.js:12` and
//! Balanced Breakfast `frontend/js/touch.js:12` carry the identical line
//!
//! ```js
//! const isTouchDevice = ('ontouchstart' in window) || (navigator.maxTouchPoints > 0);
//! ```
//!
//! and each hangs five gestures off it. Two apps, arrived at independently,
//! hand-rolling the density question this crate exists to answer.
//!
//! # The two axes are borrowed, never redefined
//!
//! Boundaries are not this crate's job. `makeover-geometry` quotes Material 3's
//! window size classes at 600 and 840 and carries [`Density`]; this crate names
//! affordances *against* those two and adds no third axis, no fourth class and
//! no breakpoint of its own. If a rule here wants a boundary that does not
//! exist, that is a conversation with `makeover-geometry`, not a constant.
//!
//! # What density is allowed to gate
//!
//! Density is a claim about **the contact patch and nothing else**. So it gates
//! affordances that depend on an interaction a fingertip cannot perform —
//! hovering, and the keyboard chrome that documents shortcuts a touch surface
//! has no way to send. It does not gate anything that is really about how much
//! screen there is. A phone is small *and* touch; a tablet is big *and* touch.
//!
//! That separation is asserted, not merely intended, by
//! `density_gates_only_what_the_contact_patch_touches`. Putting a screen-budget
//! claim on the input device is the specific failure that produced this crate,
//! and re-introducing it should have to come to the test and say so.
//!
//! # Both densities gain something (0.3.0)
//!
//! Until 0.3.0 this crate asserted that touch is pointer minus what a fingertip
//! cannot do — availability only ever subtracted as you moved from
//! [`Density::Pointer`] to [`Density::Touch`], and a test called
//! `touch_never_gains_an_affordance_pointer_lacks` said so. That claim is
//! withdrawn, deliberately, because it left a hole the crate could not name:
//! [`Affordance::Hover`]'s own doc says a fingertip has no hover state *and that
//! something else has to carry the same actions*, and nothing here was allowed
//! to be that something. [`Affordance::Anchored`] and [`Affordance::Overflow`]
//! compensate on the size axis; nothing compensated on the density axis.
//!
//! [`Affordance::Gesture`] and [`Affordance::Haptic`] are that compensation, and
//! both are gained by touch rather than lost to it.
//!
//! The claim is replaced rather than dropped. [`Affordance::gained_by`] makes
//! each member declare which density it belongs to, and
//! `a_density_member_is_available_on_exactly_the_density_it_declares` checks the
//! declaration against the rule. So a new member still cannot quietly invert:
//! it has to say which way it goes, in code, and the test is where a wrong
//! answer surfaces.
//!
//! # Collapsing is allowed, inverting is not
//!
//! Borrowed verbatim from `makeover-geometry`, where two gap relationships both
//! resolve to zero cells on a terminal and stay two members regardless. Two
//! affordances here may have identical availability today — [`Affordance::Hover`]
//! and [`Affordance::Hint`] do — and are still two members, because the call
//! site names *what is being gated*, not the rule. What must never happen is
//! one of them becoming available where the other is not for a reason that is
//! really the same reason.
//!
//! # Deliberately absent
//!
//! **A navigation shell fork.** goingson currently carries two: 12 forked
//! selectors and 10 desktop-only rules concentrated in `.app-header`, `.tab`,
//! `.tab-navigation`, `.pill-nav`, `.saved-views-sidebar` and
//! `.modal-container`. That is not one shell adapting, it is two shells, and
//! choosing to build two is a product decision rather than an adaptation. This
//! crate will not describe it, and goingson's own restructure is the way it
//! stops being true. Named here the way `makeover-layout` names validation
//! absent, so nobody has to discover it.
//!
//! **Which class applies.** The app decides, from a measured width via
//! [`SizeClass::at_width`] and from whatever it already knows about the input.
//! This crate takes both as arguments and never sniffs.
//!
//! **What a renderer does when an affordance is unavailable.** Hiding it,
//! substituting it, or showing it unconditionally anyway is renderer policy.
//! `makeover-layout` already deleted `Fill::fallback` for being exactly that.

#![forbid(unsafe_code)]

pub use makeover_geometry::{Density, SizeClass};
pub use makeover_layout::Priority;

/// An affordance whose existence depends on the surface it is offered on.
///
/// Eight members, drawn from what the two measured apps already gate by hand
/// rather than from a taxonomy. `makeover-layout`'s warning applies and is the
/// reason each addition has to point at call sites rather than at a category:
/// guessing is how a description becomes a framework.
///
/// Each answers [`Self::available`] against the two axes and nothing else. An
/// affordance that is always available is not an affordance this crate has
/// anything to say about, and `every_member_is_an_adaptation` asserts none has
/// snuck in.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub enum Affordance {
    /// Anything a consumer reveals on hover: a row's action cluster, a
    /// hover toolbar, a preview popover.
    ///
    /// Both webview apps arrived at hover-revealed row actions independently
    /// (goingson `.task-row-action`, Balanced Breakfast `.row-actions`), which
    /// is why `makeover-layout` records the reveal as behaviour of
    /// `RowPart::Actions` rather than as app policy. What neither app can say
    /// is that a fingertip has no hover state at all, so the affordance is not
    /// hidden on touch — it does not exist there, and something else has to
    /// carry the same actions.
    Hover,
    /// Chrome documenting a keyboard interaction: shortcut badges, key hints,
    /// a "press / to search" line.
    ///
    /// goingson hides `.kbd-hint` on touch. Strictly this is a claim about
    /// having a keyboard rather than about the contact patch, and [`Density`]
    /// is the closest honest proxy the family carries. Stated rather than
    /// hidden, because a detachable-keyboard tablet is where the proxy breaks
    /// and a third axis is what fixing it would cost.
    Hint,
    /// A secondary panel standing beside the primary content: a saved-views
    /// rail, a filter sidebar, an inspector.
    ///
    /// goingson hides `.saved-views-sidebar` below its widest layout. Purely a
    /// screen-budget claim — a touchscreen laptop should keep it — so this
    /// reads [`SizeClass`] alone.
    Ancillary,
    /// The detail half of a list-detail split, shown *alongside* the list
    /// rather than navigated to.
    ///
    /// goingson's `.main-content` and Balanced Breakfast's `.detail-panel`.
    /// Unavailable is not the same as absent: the detail still exists, it is
    /// reached by navigation instead of by adjacency, and which of the two a
    /// screen gets is what `makeover-layout`'s `Arrangement` is describing.
    Detail,
    /// Navigation or a primary action cluster pinned to a fixed screen edge
    /// instead of sitting in the flow of the page.
    ///
    /// The reposition bucket, and the largest single one in goingson at 20%.
    /// It exists to compensate for what a compact window cannot hold in flow,
    /// so unlike the two above it is available at the *narrow* end and not the
    /// wide one. That inversion is the point: an adaptation that only ever
    /// removes things describes a degraded layout rather than a different one.
    Anchored,
    /// An action cluster collapsed behind one control rather than laid out
    /// inline.
    ///
    /// The other compensating member. `makeover-layout`'s `Column::kept_at`
    /// already handles a *table* narrowing by dropping columns; this is the
    /// same pressure on a cluster of controls, which cannot drop any of them
    /// and folds instead.
    Overflow,
    /// A direct-manipulation gesture on content: swipe-to-action, long-press to
    /// select, pull to refresh, swipe to navigate, drag to dismiss.
    ///
    /// The density axis's compensating member, and the answer to the question
    /// [`Self::Hover`] asks and cannot answer. A fingertip has no hover state,
    /// so the row actions a pointer reveals by hovering have to arrive some
    /// other way; on both webview apps that way is a swipe.
    ///
    /// **One member for five gestures, on purpose.** goingson and Balanced
    /// Breakfast each gate all five behind a single boolean, so one member is
    /// what is measured and five would be minted from one fact. If a surface
    /// ever offers swipe without long-press, splitting this is additive and the
    /// call sites that named `Gesture` keep meaning what they meant.
    Gesture,
    /// Confirmation delivered through the contact patch rather than the eye: the
    /// tick as a drag crosses a threshold, the thump as a gesture fires.
    ///
    /// A contact-patch claim, which is what makes it this crate's business
    /// despite being the one member nothing on screen shows. goingson
    /// `frontend/js/haptics.js`.
    ///
    /// The proxy breaks where the hardware has haptics and the user or the OS
    /// has switched them off. That is neither [`Density`] nor [`SizeClass`], and
    /// it is stated here rather than fixed for the same reason [`Self::Hint`]
    /// states the detachable-keyboard case: a third axis is what fixing it would
    /// cost. **This member says the surface can, never that the user wants.**
    /// Asking the platform whether haptics are enabled is the renderer's job.
    Haptic,
}

impl Affordance {
    /// Whether this affordance exists on a surface with the given input class
    /// and screen budget.
    ///
    /// The whole crate in one call. A renderer asks per affordance and never
    /// branches on a width.
    #[must_use]
    pub const fn available(self, density: Density, size: SizeClass) -> bool {
        match self {
            // Contact patch, lost to a fingertip.
            Self::Hover | Self::Hint => matches!(density, Density::Pointer),
            // Contact patch, gained by one. A mouse can neither swipe a row nor
            // feel a confirmation, and a phone-sized window has nothing to do
            // with either.
            Self::Gesture | Self::Haptic => matches!(density, Density::Touch),
            // Screen budget. The input device has no opinion about any of them.
            Self::Ancillary => matches!(size, SizeClass::Expanded),
            Self::Detail => matches!(size, SizeClass::Medium | SizeClass::Expanded),
            Self::Anchored | Self::Overflow => matches!(size, SizeClass::Compact),
        }
    }

    /// Which [`Density`] this affordance belongs to, or `None` when it reads the
    /// screen budget instead.
    ///
    /// The replacement for the withdrawn one-directional rule (see the crate
    /// doc). Until 0.3.0 the direction was a property of the whole crate and a
    /// test enforced it globally; now each member declares its own, and
    /// `a_density_member_is_available_on_exactly_the_density_it_declares` holds
    /// the declaration to the rule.
    ///
    /// Exposed rather than kept private for the same reason [`Self::reads_density`]
    /// is: it is the crate's claim about itself, and a renderer that has one
    /// density can read it directly instead of probing [`Self::available`] twice.
    #[must_use]
    pub const fn gained_by(self) -> Option<Density> {
        match self {
            Self::Hover | Self::Hint => Some(Density::Pointer),
            Self::Gesture | Self::Haptic => Some(Density::Touch),
            Self::Ancillary | Self::Detail | Self::Anchored | Self::Overflow => None,
        }
    }

    /// Whether this affordance's availability reads [`Density`] at all.
    ///
    /// Exposed rather than kept private because it is the crate's own claim
    /// about itself: exactly the members gating a contact-patch interaction say
    /// yes. A renderer with one density can skip the rest entirely.
    #[must_use]
    pub const fn reads_density(self) -> bool {
        self.gained_by().is_some()
    }

    /// Whether this affordance's availability reads [`SizeClass`] at all.
    #[must_use]
    pub const fn reads_size(self) -> bool {
        !self.reads_density()
    }

    /// Every member, in declaration order.
    #[must_use]
    pub const fn all() -> [Self; 8] {
        [
            Self::Hover,
            Self::Hint,
            Self::Ancillary,
            Self::Detail,
            Self::Anchored,
            Self::Overflow,
            Self::Gesture,
            Self::Haptic,
        ]
    }

    /// The CSS class name an app may hang off this, without the leading dot.
    ///
    /// Present for the same reason [`SizeClass::token`] is: a webview renderer
    /// needs a stable name, and minting it per app is how two apps end up with
    /// `has-hover` and `hover-capable`.
    #[must_use]
    pub const fn token(self) -> &'static str {
        match self {
            Self::Hover => "offers-hover",
            Self::Hint => "offers-hint",
            Self::Ancillary => "offers-ancillary",
            Self::Detail => "offers-detail",
            Self::Anchored => "offers-anchored",
            Self::Overflow => "offers-overflow",
            Self::Gesture => "offers-gesture",
            Self::Haptic => "offers-haptic",
        }
    }
}

/// The column-drop cutoff a window of this size class asks for.
///
/// The seam between `makeover-layout` and `makeover-geometry` that neither
/// crate could close. Layout defines the priority ladder and `Column::kept_at`;
/// geometry defines the boundaries. Nothing said *which* cutoff a compact
/// window uses, so both webview apps answered it with `nth-child` on an ordinal
/// and inserting a column silently hid the wrong one.
///
/// A free function rather than an [`Affordance`] member because a column is not
/// gated, it is ranked: the question is which cutoff to raise to, not whether
/// the table exists.
#[must_use]
pub const fn column_cutoff(size: SizeClass) -> Priority {
    match size {
        // Only what identifies the row.
        SizeClass::Compact => Priority::Essential,
        // The optional columns go first.
        SizeClass::Medium => Priority::Secondary,
        // Everything survives.
        SizeClass::Expanded => Priority::Optional,
    }
}

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

    /// Every combination of the two axes, narrowest and coarsest first.
    fn surfaces() -> Vec<(Density, SizeClass)> {
        let mut out = Vec::new();
        for d in [Density::Pointer, Density::Touch] {
            for s in SizeClass::all() {
                out.push((d, s));
            }
        }
        out
    }

    #[test]
    fn density_gates_only_what_the_contact_patch_touches() {
        // The failure this crate exists to avoid: a screen-budget claim
        // smuggled onto the input axis, which is what let the old Touch gap
        // preset set a floor under a preset quoted from the HIG. Adding a
        // density dependency to a screen-budget affordance has to come here
        // and say so.
        for a in Affordance::all() {
            let varies_by_density = SizeClass::all()
                .iter()
                .any(|&s| a.available(Density::Pointer, s) != a.available(Density::Touch, s));
            assert_eq!(
                varies_by_density,
                a.reads_density(),
                "{a:?} disagrees with its own reads_density()"
            );
        }
    }

    #[test]
    fn size_gates_only_what_screen_budget_touches() {
        for a in Affordance::all() {
            let varies_by_size = [Density::Pointer, Density::Touch].iter().any(|&d| {
                SizeClass::all()
                    .iter()
                    .any(|&s| a.available(d, s) != a.available(d, SizeClass::Compact))
            });
            assert_eq!(
                varies_by_size,
                a.reads_size(),
                "{a:?} disagrees with its own reads_size()"
            );
        }
    }

    #[test]
    fn no_member_reads_both_axes() {
        // Not a law of adaptation, a statement about the six that exist. A
        // seventh reading both is allowed, and this test is where the claim
        // gets withdrawn rather than quietly falsified.
        for a in Affordance::all() {
            assert!(
                a.reads_density() != a.reads_size(),
                "{a:?} reads both axes; update this test and say why"
            );
        }
    }

    #[test]
    fn every_member_is_an_adaptation() {
        // A member available everywhere, or nowhere, is not describing an
        // adaptation and does not belong in this crate.
        for a in Affordance::all() {
            let yes = surfaces()
                .iter()
                .filter(|&&(d, s)| a.available(d, s))
                .count();
            assert!(yes > 0, "{a:?} exists on no surface");
            assert!(yes < surfaces().len(), "{a:?} exists on every surface");
        }
    }

    #[test]
    fn availability_is_contiguous_across_the_size_ladder() {
        // No member may exist at compact and expanded but not medium. A hole
        // in the middle is always an off-by-one, never a design.
        for a in Affordance::all() {
            for d in [Density::Pointer, Density::Touch] {
                let run: Vec<bool> = SizeClass::all()
                    .iter()
                    .map(|&s| a.available(d, s))
                    .collect();
                let transitions = run.windows(2).filter(|w| w[0] != w[1]).count();
                assert!(
                    transitions <= 1,
                    "{a:?} at {d:?} is available in a broken run: {run:?}"
                );
            }
        }
    }

    #[test]
    fn compact_compensates_rather_than_only_losing() {
        // The reposition bucket is 20% of goingson's adaptation rules and the
        // reason this crate is not just a hide-list. Whatever compact takes
        // away, something has to give back.
        for d in [Density::Pointer, Density::Touch] {
            assert!(!Affordance::Detail.available(d, SizeClass::Compact));
            assert!(Affordance::Anchored.available(d, SizeClass::Compact));
        }
    }

    #[test]
    fn hover_and_hint_collapse_and_that_is_allowed() {
        // Borrowed from makeover-geometry, where bound and peer both resolve to
        // zero cells on a terminal and stay two members. Identical rules are
        // not a duplicate; the call site names what is gated.
        for (d, s) in surfaces() {
            assert_eq!(
                Affordance::Hover.available(d, s),
                Affordance::Hint.available(d, s)
            );
        }
        assert_ne!(Affordance::Hover.token(), Affordance::Hint.token());
    }

    /// The other density. Two members, so this is total, and writing it here
    /// rather than in `makeover-geometry` keeps the axis definition borrowed
    /// rather than extended.
    fn opposite(d: Density) -> Density {
        match d {
            Density::Pointer => Density::Touch,
            Density::Touch => Density::Pointer,
        }
    }

    #[test]
    fn a_density_member_is_available_on_exactly_the_density_it_declares() {
        // Replaces `touch_never_gains_an_affordance_pointer_lacks`, withdrawn in
        // 0.3.0. That test made direction a property of the whole crate: touch
        // was pointer minus what a fingertip cannot do, and could never add.
        // Gesture and Haptic add, so the global claim had to go.
        //
        // What survives is the part worth keeping. A member still cannot invert
        // quietly -- it declares its density in `gained_by`, and this is where a
        // declaration that disagrees with the rule shows up. The old test's job
        // was to make a direction change deliberate; so is this one's.
        for a in Affordance::all() {
            match a.gained_by() {
                Some(gained) => {
                    assert!(
                        a.reads_density(),
                        "{a:?} declares a density but denies reading one"
                    );
                    for s in SizeClass::all() {
                        assert!(
                            a.available(gained, s),
                            "{a:?} declares {gained:?} but is unavailable there at {s:?}"
                        );
                        assert!(
                            !a.available(opposite(gained), s),
                            "{a:?} declares {gained:?} but is also available on the other density at {s:?}"
                        );
                    }
                }
                None => assert!(
                    !a.reads_density(),
                    "{a:?} reads density but declares no side"
                ),
            }
        }
    }

    #[test]
    fn both_densities_gain_something() {
        // The withdrawn rule, inverted into a statement of what replaced it. A
        // crate where only pointer gains members is the one this stopped being,
        // and if a refactor ever takes the touch-gained members back out, the
        // honest move is to restore the old one-directional test rather than
        // let this one quietly pass on an empty set.
        for d in [Density::Pointer, Density::Touch] {
            assert!(
                Affordance::all().iter().any(|a| a.gained_by() == Some(d)),
                "no member is gained by {d:?}"
            );
        }
    }

    #[test]
    fn the_density_axis_compensates_rather_than_only_losing() {
        // The density-axis twin of `compact_compensates_rather_than_only_losing`,
        // and the hole that motivated 0.3.0: Hover's own doc says something else
        // has to carry the actions a fingertip cannot hover to reveal, and until
        // Gesture existed nothing here could be that something.
        for s in SizeClass::all() {
            assert!(!Affordance::Hover.available(Density::Touch, s));
            assert!(Affordance::Gesture.available(Density::Touch, s));
        }
    }

    #[test]
    fn tokens_are_distinct() {
        let mut seen: Vec<&str> = Affordance::all().iter().map(|a| a.token()).collect();
        seen.sort_unstable();
        let before = seen.len();
        seen.dedup();
        assert_eq!(seen.len(), before);
    }

    #[test]
    fn the_column_cutoff_relaxes_as_the_window_widens() {
        // Priority derives Ord with Optional lowest, so a narrower window is a
        // higher cutoff. Asserted by comparison rather than by naming the three
        // constants, so reordering the ladder in makeover-layout breaks here.
        assert!(column_cutoff(SizeClass::Compact) > column_cutoff(SizeClass::Medium));
        assert!(column_cutoff(SizeClass::Medium) > column_cutoff(SizeClass::Expanded));
    }

    #[test]
    fn the_column_cutoff_replaces_the_ordinal() {
        // goingson's bug, written against this crate's answer: inserting a
        // column must not change which column drops.
        use makeover_layout::{Column, Priority as P, Width};

        let before = [
            Column {
                name: "Title",
                width: Width::Fill,
                priority: P::Essential,
            },
            Column {
                name: "Due",
                width: Width::Fixed,
                priority: P::Secondary,
            },
            Column {
                name: "Estimate",
                width: Width::Fixed,
                priority: P::Optional,
            },
        ];
        let after = [
            Column {
                name: "Title",
                width: Width::Fill,
                priority: P::Essential,
            },
            Column {
                name: "Project",
                width: Width::Fill,
                priority: P::Secondary,
            },
            Column {
                name: "Due",
                width: Width::Fixed,
                priority: P::Secondary,
            },
            Column {
                name: "Estimate",
                width: Width::Fixed,
                priority: P::Optional,
            },
        ];

        let cutoff = column_cutoff(SizeClass::Compact);
        let kept: Vec<&str> = before
            .iter()
            .filter(|c| c.kept_at(cutoff))
            .map(|c| c.name)
            .collect();
        assert_eq!(kept, ["Title"]);

        let kept: Vec<&str> = after
            .iter()
            .filter(|c| c.kept_at(cutoff))
            .map(|c| c.name)
            .collect();
        assert_eq!(kept, ["Title"]);
    }

    #[test]
    fn the_axes_are_borrowed_not_redefined() {
        // Re-exported rather than mirrored, so there is exactly one definition
        // of each in the family. A local copy is how two crates start
        // disagreeing about where 600px is.
        assert_eq!(SizeClass::Medium.min_px(), 600);
        assert_eq!(SizeClass::Expanded.min_px(), 840);
        assert_eq!(SizeClass::at_width(599), SizeClass::Compact);
    }
}