makeover-layout 0.23.0

//! The renderer-agnostic half of the make-family design system.
//!
//! <!-- wiki: makeover-layout -->
//!
//! `makeover` answers *what colour*, and varies by theme. `makeover-geometry`
//! answers *how much space*, and varies by density and surface. This crate
//! answers *what the thing is*, and varies by nothing.
//!
//! # The deferral rule
//!
//! A description names intents and relationships, never values. Say
//! [`Fill::Raised`], never `#D9DDF4`. Say `Gap::Peer`, never `6px`. What is
//! left once colour and spacing are deferred is **composition**: which edges
//! are lit, what inverts on press, what nests in what.
//!
//! The constraint that shapes all of it: a renderer that can only paint
//! rectangles has to be able to express the result. egui has no
//! `box-shadow: inset` and one stroke per widget with no per-side control; a
//! terminal has box-drawing characters and one cell of resolution, and cannot
//! draw a two-tone lit edge at all. A description that assumes per-side edges
//! is a CSS description wearing a neutral name. So this crate names the
//! *intent* — this region is a well — and each renderer chooses an expression
//! it can actually produce, including dropping half of one.
//!
//! # Scope
//!
//! Depth came first: the bevel and the surfaces it shapes. That much was
//! settled the hard way — the vocabulary here was read off audiofiles'
//! `ui::theme` and `ui::widgets`, which are the only implementation written
//! by a consumer with no CSS, then checked against both webview apps. All
//! three agreed once Balanced Breakfast's fills were corrected.
//!
//! 0.2.0 adds the rest of the description, each member drawn the same way,
//! from what the three apps already hand-write rather than from a taxonomy:
//!
//! - Components. [`Token`] (badge against chip), [`Notice`] (toast against
//!   banner), [`RowPart`], [`Heading`], [`Selector`], [`Readiness`], and
//!   [`Tone`], which is the one intent family they share.
//! - Schemas. [`Field`] for forms and [`Column`] for lists and tables.
//! - Structure. [`Region`] for the parts of a screen, [`Arrangement`] for how
//!   a screen is put together.
//!
//! **Validation** was absent on purpose here, on the grounds that neither app
//! had a shared story. That reasoning is retired — see 0.11.0 below, which is
//! where the constraints arrived and why the argument did not survive contact
//! with what the apps were measured to do.
//!
//! 0.3.0 closes a gap the first real adoption found, which is what adopting
//! against goingson first was for. [`Selector`] described only the *chosen*
//! option, so an unchosen one fell through to [`Depth::Flat`] and no renderer
//! drew it; goingson's tab strip recesses its unchosen tabs by hand and could
//! not delete the line, because being recessed is *why* the chosen tab reads as
//! coming forward. So [`Selector::unchosen`] joins `chosen`, and saying it
//! needed [`Fill::Sunken`] and [`Depth::Sunken`]: a surface set back by colour
//! with no edge, which is neither a well nor level-with.
//!
//! 0.7.0 adds [`State`], the interaction axis, closing the gap that adopting
//! against three apps rather than one made visible. The description named
//! rest and, through [`Depth::pressed`], pressed. It named neither focus nor
//! disabled, so `makeover-webview` emitted a hover rule and stopped, and each
//! consumer completed the primitive from outside by out-specifying a rule it
//! did not own: 19 such rules in goingson, 21 in the MNW server, a further set
//! in Balanced Breakfast, and three focus rings that do not match. The axis is
//! deliberately two members wide, because hover and pressed belong where they
//! already are. [`State`]'s own docs carry that argument.
//!
//! 0.8.0 finishes [`Field`], which described a field well enough to label it and
//! not well enough to draw it. Writing `makeover-webview`'s form emitter found
//! three things missing and the renderer supplied all three from outside: the
//! current value, a select's options, and the placeholder. Two of those move
//! here and one does not.
//!
//! - [`Field::placeholder`] is user-facing text sitting beside `label` and
//!   `hint`. There was never a reading on which it was renderer state; it was
//!   outside only because adding a field to a published struct is breaking.
//! - [`Field::options`] moves because every renderer needs them and each was
//!   going to invent its own shape. [`Choice`] is the shape `makeover-webview`
//!   already arrived at, taken as-is rather than redesigned.
//! - The current value stays renderer-side and is not coming here. It is the
//!   one of the three that is genuinely state: a webview reads it out of the
//!   DOM, an immediate-mode renderer holds a `&mut` to the app's own field, and
//!   a description that carried it would be a form model.
//!
//! 0.9.0 opens [`RowPart`], which was the last closed enum in the vocabulary,
//! and adds [`RowPart::Tokens`]. Both halves come from the same finding, made
//! by the first two real screens described through the router rather than by
//! reading a stylesheet.
//!
//! A goingson project card carries two trailing badges, a type and a toned
//! status; a contact card carries a primary email *and* a strip of tags. `Meta`
//! is one slot and one string, so both ports joined their facts with a
//! separator and lost what the second one was: a status reads as text where it
//! used to read as colour. [`Token`] already says exactly the right thing — a
//! small labelled thing with a kind, a tone and an optional action — and could
//! only ever be a node in its own right, never inside a row.
//!
//! So the missing thing was permission rather than a concept. `Tokens` is that
//! permission, and `#[non_exhaustive]` arrives with it so the next member is not
//! a lockstep event across three renderers. The pairing is the point: this
//! enum's own consumer in `makeover-webview` carried a comment predicting it
//! would stop compiling one day, which is a lockstep break written down and
//! waited for rather than prevented.
//!
//! Balanced Breakfast was checked before the member was added, because one
//! consumer wanting something is not evidence. It packs a count and two icon
//! buttons into the same single `Meta` slot while leaving `Actions` empty, so
//! the slot was already straining under a second consumer for a different
//! reason.
//!
//! 0.10.0 adds [`Meter`], a proportion carried as a pair rather than as a
//! percentage. Its own docs carry the argument; the short form is that the
//! percentage shape had already been tried in goingson and had already needed a
//! companion flag to recover what rounding and clamping threw away.
//!
//! 0.11.0 is four members from the quasi proving ground, batched into one
//! release because pre-1.0 a minor is breaking and a cascade is nine repos.
//! Three findings that arrived with them turned out not to belong here at all:
//! this crate has no notion of an action, a route or a destination, so anything
//! asking what a control *calls* was never the vocabulary's to say.
//!
//! - [`Figure`], a value with a caption. goingson had five of them across five
//!   screens with five class vocabularies for the one shape, which is the
//!   divergence this crate exists to end, sitting in plain sight and counted for
//!   the first time.
//! - [`RowPart::Proportion`], so a [`Meter`] can sit in a row. `Meter` reached
//!   two of its seven sites at 0.10.0 and the other five are row-shaped. Exactly
//!   [`RowPart::Tokens`]'s problem with a different payload, and it takes
//!   `Tokens`' answer: the part carries the description of a bar, not a node.
//! - [`Field::max_length`], [`Field::min`] and [`Field::max`], joining
//!   [`Field::required`], which had been sitting here as the sole constraint
//!   while the header above claimed there were none. The set stops before
//!   `pattern`, which fails the renderer test and is one site in one app.
//! - [`FieldKind::File`]. Every host has an honest answer — a native picker, an
//!   `<input type="file">`, a path prompt, an argument — and it carries no
//!   accepted-types list because `accept` appears at zero sites in either app.
//!
//! The evidence rule changed under these, and it is worth recording because four
//! earlier decisions were made under the old one. The two-app test said a shape
//! earns a word once a second app wants it. It is backwards: a rule that
//! withholds a word until a second app has duplicated the code guarantees the
//! duplication, and app three writes it a third time. The bar is now generic
//! against bespoke — is this furniture any app would have, or is it this app's
//! own? Bespoke keeps [`Region::Bespoke`], which already carries a completion
//! heatmap and is the right answer for a calendar nobody will build twice.
//!
//! 0.12.0 is two more from the same proving ground, and the same sorting
//! happened first: six findings came out of a measurement of goingson's whole
//! frontend, and four of them turned out to be asking what a control *calls*,
//! which this crate cannot say. The two that were really here:
//!
//! - [`Readiness`] grows from two states to four. It named `Ready` and
//!   `Pending` and stopped, so a screen whose list came back empty had nothing
//!   to say about it; goingson draws an empty state at 27 sites and Balanced
//!   Breakfast at 9. `Empty` and `Failed` are the same axis rather than a new
//!   member beside it, because a region shows one of the four and never two.
//!   `#[non_exhaustive]` arrives with them, the pairing [`RowPart`] made at
//!   0.9.0 and for the same reason.
//! - [`Column::sortable`], [`Column::sorted`] and [`Sort`]. The one finding in
//!   the set that completes a member rather than adding one: `Column` shipped
//!   with a width and a priority and could not say that a table is ordered by a
//!   column, so a described table could draw no caret and offer no reordering.
//!
//! 0.14.0 is two additive members on two `#[non_exhaustive]` enums, released
//! together because publishing twice for that is waste and the cascade below
//! this crate is nine repos.
//!
//! - [`Depth::Overlay`]. The enum could say raised, well, sunken and flat, and
//!   could not say that a surface sits *over* the page. Every renderer already
//!   had the surface — `makeover-tui`'s `Palette::overlay`,
//!   `makeover-immediate`'s `Palette::elevation`, `makeover-webview`'s
//!   `--elevation-overlay` — and none of them could be reached from a
//!   description. buckets_of_money has 16 modals waiting on it.
//! - [`CellPart`], which is [`RowPart`] for tables. A row's parts have carried
//!   their own content intent since 0.2.0, so `.row-actions` inherits rather
//!   than taking a text colour; a table cell had no such vocabulary and
//!   `makeover-webview` emitted one undifferentiated `.cell`, so a button in a
//!   cell was painted as text. The four members are the four things quasi's
//!   `Cell` was measured to hold, and the count is in that crate's history
//!   rather than assumed here.
//!
//! 0.15.0 adds [`FieldKind::Date`] and [`FieldKind::DateTime`], on the argument
//! [`FieldKind::Email`] was admitted on: a webview emits a different `type=`,
//! which is a native picker, the platform's validation and a different keyboard
//! on a touch device. Described as text with a "YYYY-MM-DD" hint, all three are
//! lost.
//!
//! Two members and not one or five, from a count rather than from symmetry: 13
//! sites of `date` and 13 of `datetime-local` across the MNW server and
//! goingson, and zero of `time`, `month` or `week`. The wire format each takes
//! is named here as [`DATE_FORMAT`] and [`DATETIME_FORMAT`], because a host
//! left to pick its own would disagree with a server silently, and
//! [`FieldKind::temporal`] is the pair asked about once rather than at each
//! renderer. `FieldKind`'s own comment claiming `radio` was the last HTML input
//! type missing was already false when 0.8.1 wrote it; these are what it was
//! missing.
//!
//! What each of the 0.12.0 findings deliberately leaves out is the address — what pressing a
//! header calls, and where an empty state's "Add your first project" button
//! goes. That is the boundary this crate is defined by, and four findings moved
//! across it rather than being answered here.
//!
//! 0.19.0 narrows [`State`] to [`State::Disabled`] alone. `State::Focus` is
//! gone: a description never states what has focus, because what focus *is*
//! differs per host and every renderer had already decided for itself — the
//! webview draws it from `:focus-visible`, egui refused the variant outright,
//! and quasi-tui honoured it once at startup and overrode it thereafter.
//!
//! 0.20.0 adds [`Region::Widget`], the third tier, and `#[non_exhaustive]` to
//! [`Region`] with it. Every vocabulary finding until now had two answers
//! available — grow the primitive set, or [`Region::Bespoke`] — and a whole
//! class of thing is wrong for both. A carousel is not a primitive, because a
//! terminal has none and that is the test `Node::Html` failed. It is not
//! bespoke either, because bespoke is what one app owns and every part of a
//! carousel is furniture plus members this crate already has.
//!
//! The cost of the binary was that refusing a primitive was expensive: the app
//! hand-rolls the thing forever, so the pressure always ran toward growing the
//! primitive set with one host's idioms. A named assembly changes what "no"
//! costs without changing what the vocabulary can say.
//!
//! MNW's carousel is the first consumer and was the finding that started it: one
//! partial, three pages, an ordered set of frames with a position, prev/next and
//! a dot strip, all of it sayable already and none of it nameable. See wiki
//! `widget-tier` for the ownership model, which is why this member carries a
//! name a renderer may decline to know.
//!
//! 0.21.0 adds [`Image`] and [`Fit`], found by trying to describe MNW's
//! carousel under 0.20.0's widget tier and getting one step in. Nothing named a
//! picture. The vocabulary could say a number with a caption, a badge, a meter
//! and a table, and could not say the thing three of MNW's public pages are
//! mostly made of.
//!
//! It reads as an oversight and is a measurement: 24 `<img>` sites across 22
//! MNW templates, against one in goingson and none in Balanced Breakfast or
//! audiofiles. A picture is furniture a *content platform* has, and MNW is the
//! only one in the tree, so the evidence never arrived from the two-app
//! direction the earlier rule looked in. Under the generic-against-bespoke bar
//! it is not close: a picture is not one app's own.
//!
//! A primitive rather than a widget, which is worth stating now that the tier
//! makes it a real question. A widget is an assembly of things already sayable
//! and a picture is a leaf, assembled from nothing. It also passes the test
//! `Node::Html` failed — every host has an honest answer, including a terminal,
//! which has a graphics protocol or has [`Image::alt`].
//!
//! [`Image`] carries no source, the split [`Act`] already makes: an address is
//! not this crate's to hold. See its own docs, which is where the argument is.
//!
//! 0.22.0 finishes [`Image`], which 0.21.0 shipped unable to say how much room
//! a picture needs. Without that a renderer cannot reserve space, so a picture
//! occupies nothing until its bytes arrive and then takes its full height at
//! once. Measured on MNW's landing page: a 478px jump per frame and a
//! cumulative layout shift of 0.087 for the page.
//!
//! - [`Image::intrinsic`], the picture's own dimensions, carried as [`Extent`].
//!   A fact about the asset rather than a display size, which is what keeps it
//!   on this side of the deferral rule: 5120x3412 is what the file *is*, and no
//!   renderer can learn it without fetching the bytes.
//! - [`Loading`], and the default flips to [`Loading::Eager`]. 0.21.0 emitted
//!   the webview's `loading="lazy"` for every picture, which read one
//!   consumer's habit as a rule. Deferring a picture that is on screen at first
//!   paint saves nothing and makes its shift land later. The carousel is the
//!   case that proves this cannot be one renderer-wide setting: its first frame
//!   is on screen and its others are not, in one widget, at one moment.
//!
//! 0.23.0 adds [`Showing`], which is three open findings collapsing into one
//! member. A tab group could not say which tab was open, a carousel could not
//! say which frame was up, and a disclosure could not say whether its child was
//! showing. All three are the same missing sentence, and while it was missing a
//! renderer had two moves: match on a widget name, or draw every child.
//!
//! So the widget tier was taking the blame for a gap one level below it. With
//! this a renderer derives its chrome from the description — labels get a strip,
//! no labels get previous/position/next — once, for every widget there will ever
//! be, and [`Region::Widget`]'s name goes back to being app vocabulary a
//! renderer may decline to know.
//!
//! Only the kind lives here. Which child is up, and what each child is called,
//! sit with whatever holds the regions, the same split [`Selector`] already made
//! against `Node::Select`.
//!
//! # Reach, focus and the focus ring
//!
//! Three terms, and no others, for what 0.19.0 moved out of the description.
//! **Reach** is which things can take focus and in what order; a browser reads
//! it off the document, a TUI derives it from draw order, egui from its own id
//! stack. **Focus** is which reached thing has the keyboard right now: the
//! renderer's, live, never described and never round-tripped through a
//! description. The **focus ring** is the visible cue; the token (`focus-ring`,
//! derived by `makeover` from the action colour) is the one shared artifact and
//! the drawing is the renderer's. Retired as names for any of this: "focus
//! stroke", "focus cue", "wants focus". "Caret" is a different thing — the text
//! cursor inside a field — and keeps its name.
//!
//! # Where the description stops
//!
//! A day-plan timeline, a kanban board and a calendar are not describable here
//! and will not become describable. A description expressive enough to produce
//! a timeline is a component library wearing a description's name. Generate the
//! boring 80% so the bespoke 20% gets the attention.
//!
//! [`Region::Bespoke`] is how that limit is stated rather than hidden. The
//! description names the *place* and the app owns the contents, so a screen
//! containing a timeline is still a whole screen and still routable. Without
//! it, the four goingson screens that make the app worth using would need a
//! second, undescribed path beside the router, and two paths is how a
//! vocabulary starts drifting from its app again.
//!
//! [`Region::Widget`] sits between that limit and the primitives, and it does
//! not move the limit. A widget is an assembly of members this crate *already*
//! has, under a name a renderer may or may not recognise. Anything that needs a
//! member the vocabulary does not have is still a finding about the vocabulary
//! or still bespoke; naming an assembly buys no new expressive power, which is
//! exactly why it is safe to let the set grow outside this crate.

#![forbid(unsafe_code)]

/// A colour intent this crate refers to but never resolves.
///
/// The string is the token name `makeover` publishes, so a renderer can look
/// it up without this crate knowing what colour came back.
pub trait Intent {
    /// The `makeover` intent token this resolves against.
    fn token(self) -> &'static str;
}

/// Which way the light falls across a two-tone edge.
///
/// The whole content of a bevel, once colour and thickness are deferred. The
/// light is always assumed to come from the top left: every consumer measured
/// agreed on that and none of them ever varied it, so it is an invariant here
/// rather than a parameter.
///
/// # The two corners that belong to both edges
///
/// Top-right and bottom-left are where the lit run meets the shaded one, and
/// the description's claim is that they belong to *both*. How a renderer says
/// that is its own business, because the answer is bounded by resolution and
/// not by taste:
///
/// - A terminal cell is roughly 8x17 device pixels, so giving the whole corner
///   to one tone thickens that edge by a cell and reads as one run overrunning
///   the other. A half-cell glyph divides the cell already, so `makeover-tui`
///   splits it and recovers real information. Its box-drawing fallback cannot:
///   a single stroke has no half to give, so there both corners go to dark.
/// - A pixel bevel is a one-point stroke by default, which makes the corner a
///   one-point square. There is nothing to divide — a diagonal seam across one
///   point is sub-pixel, and antialiasing renders it as the blend a mitred join
///   already produces. So `makeover-immediate` mitres and is *not* diverging;
///   it is the same rule at a resolution where the split degenerates.
///
/// Stated here so the difference reads as a decision rather than as drift. A
/// renderer with room to divide the corner should; one without should mitre or
/// pick the shaded tone, and neither is a bug.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum Bevel {
    /// Lit from the top left: light on top and left, dark on bottom and right.
    Raised,
    /// The same edge inverted, which is also the pressed state of anything
    /// that draws itself [`Bevel::Raised`].
    Inset,
}

impl Bevel {
    /// The edge intents, as `(top_left, bottom_right)`.
    ///
    /// Split out from any painting because the inversion *is* the idea, and
    /// it is the one part every renderer implements identically.
    #[must_use]
    pub const fn edges(self) -> (Edge, Edge) {
        match self {
            Self::Raised => (Edge::Light, Edge::Dark),
            Self::Inset => (Edge::Dark, Edge::Light),
        }
    }

    /// Pressing inverts. A raised control reads as inset while held.
    ///
    /// Stated here rather than left to each consumer because a cascade can
    /// carry a pressed state and an immediate-mode renderer cannot: audiofiles
    /// resolves this per call site, eighteen times.
    #[must_use]
    pub const fn pressed(self) -> Self {
        match self {
            Self::Raised => Self::Inset,
            Self::Inset => Self::Raised,
        }
    }
}

/// One side of a bevel, named by the intent it takes.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum Edge {
    /// The lit side.
    Light,
    /// The shadowed side.
    Dark,
}

impl Intent for Edge {
    fn token(self) -> &'static str {
        match self {
            Self::Light => "bevel-light",
            Self::Dark => "bevel-dark",
        }
    }
}

/// A surface intent a region is filled with.
///
/// `#[non_exhaustive]`, so a renderer must carry a wildcard arm and a new
/// member is additive rather than breaking. Added 0.4.0, after [`Sunken`]
/// (an additive member, 0.3.0) hard-broke `makeover-tui` and
/// `makeover-immediate` at compile time and left neither able to move until
/// both published. The vocabulary exists to grow and the renderers exist to
/// disagree about how much of it they answer, so growth must not be a
/// lockstep event. The renderer's wildcard is not a hole: [`Fill`] is
/// resolved through a fallible lookup, and a missing intent is answered with
/// structure rather than with a substituted colour.
///
/// [`Sunken`]: Fill::Sunken
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub enum Fill {
    /// The page behind everything.
    Page,
    /// A surface lifted off the page: cards, controls, menus, toasts.
    Raised,
    /// A surface floating above the page rather than resting on it.
    Overlay,
    /// The inside of a well.
    Well,
    /// A surface set back from the one it sits on, by colour and nothing else.
    ///
    /// Not a well. A well is a hole with an edge, and the two are authored in
    /// opposite directions: `makeover` derives `surface-well` by inverting
    /// against the theme's own content colour, while `surface-sunken` is
    /// authored and free to sit darker than raised (goingson's does). Naming
    /// only the well left the recessed-with-no-edge surface unsayable, which is
    /// what an unchosen tab is: it recedes so the chosen one can come forward,
    /// and it carries no bevel of its own.
    ///
    /// Added 0.3.0, from goingson's tab strip, which hand-writes exactly this
    /// and could not delete the line because no member described it.
    Sunken,
}

// No `fallback` here, deliberately. An earlier cut had `Fill::Well` fall back
// to `Fill::Page` so a consumer on makeover 2.2.0, which has no `surface-well`,
// had something to paint. makeover-tui found that wrong within a day: page is
// the surface a well is usually cut into, so on a terminal that substitution
// produces exactly the invisibility it was meant to prevent, and the right
// answer there is a drawn edge rather than a different colour.
//
// Substituting one intent for another is renderer policy. The description says
// what the region is and stops.

impl Intent for Fill {
    fn token(self) -> &'static str {
        match self {
            Self::Page => "surface-page",
            Self::Raised => "surface-raised",
            Self::Overlay => "surface-overlay",
            Self::Well => "surface-well",
            Self::Sunken => "surface-sunken",
        }
    }
}

/// How a region sits relative to the surface behind it.
///
/// Fill and bevel are named together because naming them apart is what let
/// them disagree. Every consumer measured had at least one region carrying a
/// raised bevel over a recessed fill: audiofiles fixed it in `raised_frame`
/// and recorded the bug in its doc comment, and Balanced Breakfast still had
/// twelve of them a year later. A single name for the pair makes that
/// unrepresentable.
/// `#[non_exhaustive]` for the same reason as [`Fill`], and in the same
/// release: a depth this renderer has no drawing for should cost it a
/// wildcard arm, not a compile error and a wait on someone else's publish.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub enum Depth {
    /// Level with its surroundings. No edge.
    Flat,
    /// A card laid on the panel it sits in.
    Raised,
    /// A hole in the panel, with content down inside it. For anything the
    /// user looks *into*: a table body, a tag tree, a text field.
    Well,
    /// Set back from what it sits on, by colour alone. No edge.
    ///
    /// The one member carrying a fill without a bevel, so a renderer cannot
    /// assume the two arrive together. That is deliberate and it is still the
    /// pairing rule: both halves come off the same `Depth`, so they cannot
    /// disagree, and here one half is legitimately absent.
    ///
    /// Distinct from [`Depth::Flat`], which has no fill either and inherits.
    /// Recessed and level-with are different claims, and only one of them
    /// needs a colour.
    Sunken,
    /// A surface sitting *over* the page rather than in it. A modal, a popover,
    /// a menu.
    ///
    /// Takes elevation and no bevel: a surface overlaying the page is lifted
    /// off it, and a surface in the page is cut into it. That is the same
    /// pairing rule the rest of the enum holds, applied to the one case where
    /// the separation is not an edge at all — the lift and the scrim behind it
    /// are already saying where the surface is.
    ///
    /// Every renderer had the surface before it had this variant.
    /// `makeover-tui` carries `Palette::overlay`, `makeover-immediate` gained
    /// `Palette::elevation` at 0.10.0, and `makeover-webview` emits
    /// `--elevation-overlay`. What was missing was the route from a description
    /// to any of them, which is why this is one variant rather than a feature.
    Overlay,
}

impl Depth {
    /// The edge this depth is drawn with, if it has one.
    #[must_use]
    pub const fn bevel(self) -> Option<Bevel> {
        match self {
            // Sunken joins Flat here, for the opposite reason: Flat has no edge
            // because nothing separates it from its surroundings, and Sunken has
            // none because its colour is already doing the separating.
            Self::Flat | Self::Sunken => None,
            // A third reason to have no edge, which is why it gets its own arm
            // rather than joining the two above: an overlay is separated by the
            // lift and by the scrim behind it, so an edge would be a second
            // answer to a question already answered.
            Self::Overlay => None,
            Self::Raised => Some(Bevel::Raised),
            Self::Well => Some(Bevel::Inset),
        }
    }

    /// The surface this depth is filled with.
    ///
    /// [`Depth::Flat`] has no fill of its own: it inherits whatever it sits on,
    /// which is the difference between level-with and painted-the-same-colour.
    #[must_use]
    pub const fn fill(self) -> Option<Fill> {
        match self {
            Self::Flat => None,
            Self::Raised => Some(Fill::Raised),
            Self::Well => Some(Fill::Well),
            Self::Sunken => Some(Fill::Sunken),
            Self::Overlay => Some(Fill::Overlay),
        }
    }

    /// Pressing a raised region reads as a well, and nothing else moves.
    ///
    /// [`Depth::Overlay`] is untouched along with the rest: an overlay is a
    /// surface, not a control, so there is nothing there to press.
    #[must_use]
    pub const fn pressed(self) -> Self {
        match self {
            Self::Raised => Self::Well,
            other => other,
        }
    }
}

/// An interaction state a region can be in, beside whatever [`Depth`] it is.
///
/// Orthogonal to depth on purpose. A disabled button is still [`Depth::Raised`]
/// and a disabled field is still a [`Depth::Well`], so folding either member
/// into `Depth` would make [`Depth::bevel`] and [`Depth::fill`] answer for
/// something that is not a depth, and would leave disabled-button and
/// disabled-field sharing one variant that cannot tell them apart.
///
/// # Why hover and pressed are not members
///
/// The line is whether every renderer has the state to express, not whether CSS
/// does. Hover is renderer policy and `makeover-webview` says so in its own
/// header: a terminal and an immediate-mode painter have no pointer hovering
/// over anything, and pressed already arrives through [`Bevel::pressed`] and
/// [`Depth::pressed`], where it belongs, because pressing is a depth inversion
/// rather than a separate condition.
///
/// Focus and disabled are different in kind. A TUI has a focused widget and a
/// greyed-out one; so does egui. Both were unsayable here, so all three webview
/// consumers supplied them from outside the primitive by out-specifying rules
/// they did not own: goingson alone carries 19 of them, and the MNW server
/// another 21. That is the divergence this crate exists to end, arriving one
/// layer down.
///
/// # The principle this encodes
///
/// A primitive owns every state it implies. A renderer that emits a hover rule
/// for a thing owes disabled and the capability answer for that same thing,
/// because anything less exports the completion work to N consumers who will
/// each do it differently.
///
/// Focus is not on that list and was removed from this axis in 0.19.0. It is
/// the renderer's, decided after the description; see the crate header, "Reach,
/// focus and the focus ring", for the three terms and who owns each.
///
/// `#[non_exhaustive]` for the reason [`Fill`] and [`Depth`] carry it: growth
/// must not be a lockstep event across the three renderers.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub enum State {
    /// Present, visible, and not answering.
    ///
    /// Not the same as absent, and deliberately not a [`Fill`]: a disabled
    /// control keeps the surface it always had and stops responding, so what
    /// changes is its content and its interactivity rather than what it is.
    Disabled,
}

impl State {
    /// Whether a region in this state stops answering the pointer.
    ///
    /// Stated in the description rather than left to each renderer, on the same
    /// reasoning as [`Bevel::pressed`]: a cascade carries it for free and an
    /// immediate-mode renderer resolves it per call site, so leaving it unsaid
    /// means resolving it once per consumer and disagreeing.
    #[must_use]
    pub const fn suppresses_interaction(self) -> bool {
        // A match rather than a bare `true`, so a member added to this
        // `#[non_exhaustive]` axis has to answer the question rather than
        // inheriting an answer.
        match self {
            Self::Disabled => true,
        }
    }
}

impl Intent for State {
    fn token(self) -> &'static str {
        match self {
            // Reusing the muted content intent rather than minting a
            // `disabled` colour. Disabled is a reduction and not a status, and
            // `makeover-webview`'s progress rules already record the reading
            // that `content-muted` is what disabled looks like.
            Self::Disabled => "content-muted",
        }
    }
}

/// What a region is saying, when it is saying something.
///
/// The one intent family shared by badges, notices and nothing else. Kept
/// separate from [`Fill`] because a surface is where a thing sits and a tone is
/// what it means, and the three apps agree on the four statuses:
/// `info_banner` / `warning_banner` in audiofiles, `.toast-info` /
/// `.toast-success` / `.toast-error` in goingson, `.toast.success` /
/// `.toast.error` in Balanced Breakfast.
///
/// The per-tag palette (`category-one` through `category-six`) is deliberately
/// not here. Which colour a *particular* tag takes is app domain, and both
/// webview apps already carry it as a `data-color` attribute.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum Tone {
    /// No status. Reads as ordinary de-emphasised content.
    Neutral,
    /// Something worth knowing and nothing to do about it.
    Info,
    /// Something finished and it worked.
    Success,
    /// Something the user should look at before continuing.
    Warning,
    /// Something broken, or something about to be destroyed.
    Danger,
}

impl Intent for Tone {
    fn token(self) -> &'static str {
        match self {
            // Neutral has no status token of its own. It takes the muted
            // content intent, which is what both webview apps already spell as
            // `data-color="muted"`.
            Self::Neutral => "content-muted",
            Self::Info => "info",
            Self::Success => "success",
            Self::Warning => "warning",
            Self::Danger => "danger",
        }
    }
}

/// A small labelled thing that sits inside something else.
///
/// Two members, because the three apps drew three taxonomies and only one line
/// runs through all of them: does it answer a click. audiofiles has
/// `classification_badge` (a label) against `tag_chip`, `tag_chip_removable`
/// and `selectable_tag` (all of which do). Balanced Breakfast has `.tag` and
/// `.badge` against `.tag-chip`. goingson is the one that has to move: its
/// `.tag` and `.badge` are a single CSS rule, so every call site has to be read
/// to decide which of the two it always was.
///
/// The evidence that a chip is a real concept rather than a badge with a
/// cursor: audiofiles inverts its bevel on press and Balanced Breakfast latches
/// `.tag-chip.active` with the inset bevel. Two independent arrivals at "a chip
/// holds itself down", which is exactly what [`Depth::pressed`] already says.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum Token {
    /// Non-interactive status or count. Answers no click.
    Badge,
    /// An interactive or removable token. Answers a click, and latches if it
    /// stands for a filter that is either on or off.
    Chip {
        /// Whether it carries its own remove affordance.
        removable: bool,
    },
}

impl Token {
    /// Whether this answers a click.
    ///
    /// The whole difference between the two members, and the reason a renderer
    /// with no hover (a touch surface, a terminal) can still tell them apart.
    #[must_use]
    pub const fn interactive(self) -> bool {
        matches!(self, Self::Chip { .. })
    }

    /// How it sits, given whether it is currently latched down.
    ///
    /// A badge is flat: it is a label, and giving it an edge would say it can
    /// be pressed. A chip is raised, and inset while latched.
    #[must_use]
    pub const fn depth(self, latched: bool) -> Depth {
        match self {
            Self::Badge => Depth::Flat,
            Self::Chip { .. } if latched => Depth::Well,
            Self::Chip { .. } => Depth::Raised,
        }
    }
}

/// Something the app is telling the user, unprompted.
///
/// Two concepts, not one with a placement. They differ in more than where they
/// sit: a toast is transient, stacked and self-dismissing, and a banner is
/// persistent, in flow, one per region, and dismissed by fixing the condition
/// it reports. Folding them into one member with a placement parameter would
/// make lifetime, stacking and dismissal all placement-dependent, which is the
/// description leaking renderer policy.
///
/// All three apps have banners: `info_banner` and `warning_banner` in
/// audiofiles, five of them in goingson (sync, sync-result, vacation-day,
/// timer-active, past-review), `.update-banner` in Balanced Breakfast. The two
/// webview apps also have toasts. So neither member is speculative, and no app
/// gains a concept it lacks except audiofiles, whose renderer may legitimately
/// decline to draw a toast at all.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum Notice {
    /// Transient, stacked, dismisses itself.
    Toast,
    /// Persistent, in flow, one per region, dismissed by fixing the cause.
    Banner,
}

impl Notice {
    /// Whether it goes away on its own.
    #[must_use]
    pub const fn transient(self) -> bool {
        matches!(self, Self::Toast)
    }

    /// How it sits.
    ///
    /// A toast floats above the page rather than resting on it, which is
    /// [`Fill::Overlay`]'s whole reason to exist. A banner is a card in the
    /// flow. Both are raised, and they are raised off different things.
    #[must_use]
    pub const fn fill(self) -> Fill {
        match self {
            Self::Toast => Fill::Overlay,
            Self::Banner => Fill::Raised,
        }
    }
}

/// The parts of a list row.
///
/// Four to begin with, taken from Balanced Breakfast, which was the only
/// consumer that had all of them (`row-primary`, `row-secondary`, `row-meta`,
/// `row-actions`). audiofiles has two and no slot structure at all, so it gains
/// meta and actions as real work rather than a rename; goingson moves off
/// `task-row` / `task-cell`.
///
/// [`Tokens`](Self::Tokens) joined at 0.9.0, and `#[non_exhaustive]` with it.
/// See the crate header for why the two arrived together.
///
/// # Meta against Tokens
///
/// The line is whether the thing has its own standing. `Meta` is one short
/// trailing fact about the row, written as text: a count, a size, a date.
/// `Tokens` is a set of small labelled things, each of which can be toned and
/// can answer a click. "3 files" is meta. A status badge that is amber, and a
/// tag you can click to filter by, are tokens.
///
/// Keeping them apart is what a single widened slot would have foreclosed. A
/// renderer can right-align one string and cannot usefully do the same to a
/// strip of chips, and a fact that is not clickable should not be drawn as
/// though it were.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub enum RowPart {
    /// The thing itself. What the row is called.
    Primary,
    /// Supporting text under the primary.
    Secondary,
    /// A short trailing fact: a count, a size, a date.
    Meta,
    /// Controls that act on this row.
    Actions,
    /// Small labelled things belonging to the row: badges, chips, tags.
    ///
    /// Each carries its own [`Token`] kind and [`Tone`], so a renderer with no
    /// colour still has the kind to work with, and one with no chips still has
    /// the label. That is the constrained-consumer test this vocabulary exists
    /// to pass, and it is why the tone lives on the token rather than on the
    /// part.
    Tokens,
    /// How much of a set the row's thing has done: a [`Meter`] in the row.
    ///
    /// Added 0.11.0, `da5666ae`, and it is [`Tokens`](Self::Tokens)'s problem
    /// again with a different payload. [`Meter`] arrived at 0.10.0 and closed
    /// two of the seven sites that asked for it; the other five sit in rows, and
    /// a row holds no nodes by the ruling that a row part may not carry an
    /// arbitrary node — the door through which a description becomes a
    /// templating language. So the part carries the *description of a bar*
    /// rather than a node, exactly as `Tokens` carries tags rather than nodes.
    ///
    /// Without it a row flattens the proportion into [`Meta`](Self::Meta) as
    /// "3/7 subtasks", which keeps both numbers and loses the reading, the same
    /// way a toned status badge read as prose before `Tokens`.
    Proportion,
}

impl RowPart {
    /// The content intent the part takes.
    #[must_use]
    pub const fn intent(self) -> &'static str {
        match self {
            Self::Primary => "content",
            Self::Secondary => "content-secondary",
            Self::Meta => "content-muted",
            // Actions carry controls rather than text, so they inherit.
            Self::Actions => "content",
            // So do tokens: each one carries its own tone, and a part-level
            // intent underneath it would fight the token that sits on it.
            Self::Tokens => "content",
            // And so does a proportion, for the same reason: the meter carries
            // the tone, and it is about the ratio rather than about the row.
            Self::Proportion => "content",
        }
    }
}

/// How far down the heading tree a title sits.
///
/// Three, and only the three that are actually headings. The bands those used
/// to be filed with (goingson's `.page-header`, Balanced Breakfast's `.header`
/// and `.detail-header`) are arrangement, not type, and live at
/// [`Region::Band`]. One of them contains no text at all.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum Heading {
    /// Names the whole screen. One per screen.
    Page,
    /// Names a block within the screen.
    Section,
    /// Names a sub-block inside an already-named section.
    Subsection,
}

impl Heading {
    /// Whether a rule follows the heading.
    ///
    /// audiofiles' `section_header` draws a separator and its
    /// `subsection_label` deliberately does not, which is the only thing
    /// distinguishing the two once weight and colour are deferred.
    #[must_use]
    pub const fn separated(self) -> bool {
        matches!(self, Self::Section)
    }
}

/// A control that picks between things.
///
/// Three, because three distinct behaviours are in play and collapsing any two
/// loses something. A segmented control picks a value; a tab picks a pane; a
/// toggle picks nothing and simply holds itself on or off.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum Selector {
    /// Exactly one of N, and the options abut.
    Segmented,
    /// Independent on or off, on its own.
    Toggle,
    /// Navigation between panes. The folder semantic.
    Tabs,
}

impl Selector {
    /// How the chosen option sits.
    ///
    /// Held in for a segmented control and a toggle, which is the same shape
    /// pressing produces and the whole economy of the idiom: one appearance,
    /// two reasons to wear it. A tab is the exception, because the selected
    /// folder tab comes *forward* to join the pane it opens.
    #[must_use]
    pub const fn chosen(self) -> Depth {
        match self {
            Self::Segmented | Self::Toggle => Depth::Well,
            Self::Tabs => Depth::Raised,
        }
    }

    /// How the options that were *not* picked sit.
    ///
    /// Added 0.3.0. Describing only [`Selector::chosen`] left the unchosen
    /// option falling through to [`Depth::Flat`], which says it is level with
    /// the strip it sits in, and no renderer emitted anything for it. That is
    /// wrong in both directions and goingson proved it: its unchosen tabs are
    /// recessed by hand, and being recessed is *why* the chosen one reads as
    /// coming forward. Against a flat strip, a raised chosen tab is a bevel
    /// drawn on the strip's own colour, which is a much weaker folder effect
    /// than the contrast the idiom is named after.
    ///
    /// Each member is the inverse of its chosen state, which is the whole
    /// content of "picked" once colour is deferred:
    ///
    /// - Tabs recede, so the chosen one comes forward.
    /// - A segment and a toggle stand up, so the chosen one is held in.
    #[must_use]
    pub const fn unchosen(self) -> Depth {
        match self {
            Self::Tabs => Depth::Sunken,
            Self::Segmented | Self::Toggle => Depth::Raised,
        }
    }

    /// Whether the options touch.
    ///
    /// The gap is the entire difference between a segmented control and a row
    /// of buttons that happen to sit near each other, which is what audiofiles'
    /// `segmented_control` says in its own comment and why it zeroes the
    /// spacing by hand.
    #[must_use]
    pub const fn abutting(self) -> bool {
        matches!(self, Self::Segmented | Self::Tabs)
    }
}

/// What is in a region right now.
///
/// The state, not the shimmer. Whether pending paints a skeleton, a spinner or
/// nothing at all is renderer policy, the same class of decision that got
/// `Fill::fallback` deleted from this crate. goingson and Balanced Breakfast
/// each grew a skeleton with differently-named parts; both keep them, as the
/// webview renderer's expression of [`Readiness::Pending`]. audiofiles has none
/// and needs none, because an immediate-mode renderer simply repaints.
///
/// # Four states and not two, as of 0.12.0
///
/// `703f4cd2`. It named `Ready` and `Pending` and stopped, so a described screen
/// whose list came back empty had to render an empty region or invent its own
/// placeholder text, and neither says what it is. goingson draws one at 27 sites
/// across 12 files and Balanced Breakfast at 9, with a class family that had
/// already drifted into `empty-state`, `empty-state--error`, `error-state` and
/// six more.
///
/// The four are one axis because they are mutually exclusive: a region shows its
/// content, or a sign that it is coming, or a sign that there is none, or a sign
/// that it broke. Never two. That is the test for one enum against several
/// fields, and it is why this grew rather than a new member arriving beside it.
///
/// # What is not here
///
/// **The message.** "No projects yet" is content, and this names a state. It
/// lives with whatever holds the region — in quasi's case a `Slot` — alongside
/// the action that leads out of the emptiness, since an address is the one thing
/// this crate never names.
///
/// **How much room it gets.** goingson's `--compact`, `--dashboard` and
/// `--padded` are the same state at three sizes, and a size is
/// `makeover-geometry`'s question. Naming them here would be this crate stating
/// values again.
///
/// **The icon.** Presentation, and each host has its own answer or none.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub enum Readiness {
    /// The content is here.
    Ready,
    /// The content is on its way.
    Pending,
    /// The content arrived and there is none of it.
    ///
    /// Not a failure. An empty list is the normal state of a new install, and a
    /// renderer that drew it in a danger tone would be reporting a fault where
    /// there is none.
    Empty,
    /// The content did not arrive.
    Failed,
}

impl Readiness {
    /// Whether the region draws its own content, or something standing in for
    /// it.
    ///
    /// The question every renderer asks first, so it is answered once here
    /// rather than by a `matches!` in each. A state added later is a stand-in
    /// until proven otherwise: falling back to drawing content that may not be
    /// there is the worse of the two mistakes.
    #[must_use]
    pub const fn shows_content(self) -> bool {
        matches!(self, Self::Ready)
    }

    /// What the state means, for a renderer choosing a colour.
    ///
    /// Derived rather than carried, which is the opposite of [`Meter`] and
    /// [`Figure`], and the difference is worth stating: a proportion's meaning
    /// depends on what is being counted and only the app knows it, while
    /// "nothing here yet" and "this broke" mean the same thing in every app that
    /// will ever have them.
    #[must_use]
    pub const fn tone(self) -> Tone {
        match self {
            Self::Failed => Tone::Danger,
            _ => Tone::Neutral,
        }
    }
}

/// How much of a set is done.
///
/// Added 0.10.0. Nine sites across the two webview apps drew a bar and nothing
/// here named one, so every described screen concatenated the two numbers into
/// its heading text instead: "Subtasks 3/7", "Time Tracking 45m tracked / 30m
/// est, over". Every fact survives that and the reading does not, which is the
/// same loss `RowPart::Tokens` closed when a toned status badge became prose.
///
/// # Why a pair and not a percentage
///
/// Both numbers, not the percentage the apps compute from them. The percentage
/// was the obvious shape and it had already been tried: goingson's
/// `Task::time_progress` divides, rounds, and then clamps to 100, which throws
/// away the one case the bar exists to show — 45 minutes tracked against a
/// 30-minute estimate. It carries a separate `is_over_estimate` boolean beside
/// it to recover the fact the clamp dropped. A pair keeps the over-run without a
/// companion flag, and [`percent`](Meter::percent) is still one call away for a
/// renderer that wants it.
///
/// The pair is also what the apps already have at every site. All seven
/// determinate bars write the ratio into the accessible layer and never the
/// percentage: `title="3/7 subtasks"`, `aria-label="3 of 7 subtasks completed"`,
/// a milestone's own `3/7` span. Given 43 nothing can recover "3 of 7", so a
/// percentage member would have made [`label`](Meter::label) mandatory at every
/// call site, which is the concatenated text this member removes, moved one
/// layer down.
///
/// # What this is not
///
/// The progress of an *operation*. Two of the nine sites are that — goingson's
/// focus timer, Balanced Breakfast's feed fetch — and they get nothing here, on
/// purpose. Both are imperative controllers over a live handle, driven by a tick
/// or an event stream, and a description is built once and dropped. Holding one
/// would mean growing a way to update a description between renders, which is a
/// different feature. [`Readiness::Pending`] and a [`Notice::Toast`] carry the
/// honest part.
///
/// The two cases are distinguishable in the markup rather than by taste: every
/// determinate bar in both apps carries a tone, and neither operation bar
/// carries one. Two codebases drew that line the same way without coordinating.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub struct Meter<'a> {
    /// How much is done. May exceed [`total`](Self::total), and that is the
    /// case worth drawing.
    pub done: u32,
    /// How much there is to do. Zero means there is no set, not that the set is
    /// complete.
    pub total: u32,
    /// What the proportion means right now.
    ///
    /// Carried rather than derived, because no renderer can work it out. The
    /// same 90% is [`Tone::Success`] on a subtask rollup and [`Tone::Danger`] on
    /// a time estimate, and goingson picks between them from `is_over_estimate`,
    /// a fact about the data and not about the number.
    pub tone: Tone,
    /// What is being counted, if the bar says so: "subtasks", "tasks".
    ///
    /// The noun, not the ratio. A renderer builds "3 of 7 subtasks" from this
    /// and the two numbers; handing it the assembled string would put the
    /// sentence order in the description, where a terminal at one line and a
    /// tooltip want different ones.
    pub label: Option<&'a str>,
}

impl<'a> Meter<'a> {
    /// A proportion with no tone and no label.
    #[must_use]
    pub const fn new(done: u32, total: u32) -> Self {
        Self {
            done,
            total,
            tone: Tone::Neutral,
            label: None,
        }
    }

    /// What the proportion means.
    #[must_use]
    pub const fn tone(mut self, tone: Tone) -> Self {
        self.tone = tone;
        self
    }

    /// What is being counted.
    #[must_use]
    pub const fn label(mut self, label: &'a str) -> Self {
        self.label = Some(label);
        self
    }

    /// How full the bar is, 0 to 100, clamped.
    ///
    /// For drawing, which is the only thing a clamped number is good for. Ask
    /// [`overflowing`](Self::overflowing) before reporting it as a fact, or this
    /// is `time_progress`'s bug again with the clamp moved.
    ///
    /// An empty set reads as 0. Nothing is done, because there is nothing to do
    /// and no bar to fill; the apps guard on the count before drawing at all.
    #[must_use]
    pub const fn percent(&self) -> u8 {
        if self.total == 0 {
            return 0;
        }
        let scaled = (self.done as u64 * 100) / self.total as u64;
        if scaled > 100 { 100 } else { scaled as u8 }
    }

    /// Whether more is done than there was to do.
    ///
    /// The fact [`percent`](Self::percent) destroys, kept reachable so a
    /// renderer can mark the over-run rather than drawing a full bar and
    /// implying it landed exactly.
    #[must_use]
    pub const fn overflowing(&self) -> bool {
        self.done > self.total
    }

    /// Whether there is a set at all.
    ///
    /// A meter over nothing is sayable on purpose, for the same reason a field
    /// with no options is: it is what an app with an unloaded count actually
    /// has, and a renderer that shows an empty bar says so on screen rather than
    /// dividing by zero.
    #[must_use]
    pub const fn is_empty(&self) -> bool {
        self.total == 0
    }
}

/// One figure with a caption: a number and what it counts.
///
/// The dashboard shape. A large value over a small caption, several of them in a
/// strip: a current streak, a completion rate, a total. Added 0.11.0,
/// `93c6a174`, after goingson turned out to have five of them across five
/// screens with five class vocabularies for the one shape — `task-overview-stat`,
/// `stat-box`, `month-stat-item`, `contact-summary-stat`, `sync-stat`. Four put
/// the value above the caption and one inverts it, which is drift inside the
/// shape rather than a second shape.
///
/// # Why the value is text
///
/// "17", "84%", "12/30", "3d". A figure is whatever the app computed, already
/// formatted, and the formatting is the app's because only it knows whether the
/// number is a percentage, a duration or a ratio. This carries none of the
/// arithmetic [`Meter`] carries, and that is the difference between them: a
/// meter is a proportion a renderer draws, and a figure is a fact a renderer
/// sets in type.
///
/// # Tone is carried, for [`Meter`]'s reason
///
/// Three of the five sites tone the figure by their own means — `red`/`blue` on
/// the weekly review, a `${type}` class on the monthly one, `sync-stat-warn` on
/// sync. So tone is carried at every site that needs it and derived at none, and
/// no renderer can work out that a streak of zero is worth colouring.
///
/// # What is not here
///
/// Whether the figure answers a click. One of the five is a control — sync's
/// "Not Applied: 3" opens the list — and an action is not something this crate
/// can name: nothing here knows what a route is. That belongs beside the figure
/// in whatever layer holds the actions, the same way a row's activation sits
/// beside its parts rather than inside them.
///
/// The arrangement is not here either. Several figures in a strip is a set, and
/// a renderer given them one at a time cannot tell it is looking at one; the
/// layer that holds the tree is where the set gets said.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub struct Figure<'a> {
    /// The number, formatted the way the app means it to read.
    pub value: &'a str,
    /// What it counts. The caption under the value.
    pub caption: &'a str,
    /// How the value has moved, if the app is tracking that.
    ///
    /// Added 0.13.0. Text, for [`value`](Self::value)'s reason: only the app
    /// knows whether a move reads as `+12.5%`, `+3` or `2x`, and a renderer
    /// handed a number would have to guess.
    ///
    /// This is what [`tone`](Self::tone) was for and had no consumer of. The MNW
    /// server has four screens whose stat card is a label, a value and a delta,
    /// and the delta is the toned part: the figure itself is an ordinary fact
    /// and it is the movement that reads as good or bad. Without this the delta
    /// has to be folded into the caption, which loses the tone and reads as a
    /// longer caption rather than as a second, smaller line.
    pub change: Option<&'a str>,
    /// What the figure means right now. [`Tone::Neutral`] is an ordinary fact.
    ///
    /// Applies to [`change`](Self::change) where there is one, since that is the
    /// part that carries the judgement, and to the value where there is not.
    pub tone: Tone,
}

impl<'a> Figure<'a> {
    /// A figure that is an ordinary fact.
    #[must_use]
    pub const fn new(value: &'a str, caption: &'a str) -> Self {
        Self {
            value,
            caption,
            change: None,
            tone: Tone::Neutral,
        }
    }

    /// How the value has moved.
    #[must_use]
    pub const fn change(mut self, change: &'a str) -> Self {
        self.change = Some(change);
        self
    }

    /// What the figure means.
    #[must_use]
    pub const fn tone(mut self, tone: Tone) -> Self {
        self.tone = tone;
        self
    }
}

/// Something the user can do, and what it costs to say so.
///
/// Added 0.17.0, out of `quasi-tui`: the terminal renderer had drawn one of
/// these for months and every other consumer that wanted a button had written
/// its own, because this layer named [`RowPart::Actions`] as a *slot* and never
/// named the thing that goes in it. Beside [`Meter`] and [`Figure`] for the
/// reason those are here: a renderer that is handed the parts has to decide how
/// to say them, and a renderer that is handed a finished string has already had
/// the decision made for it.
///
/// No address. Where a control goes is the app's business and every host
/// follows it differently — an `hx-get`, a protocol URL, a function call — so
/// the description says what the control *is* and the caller keeps what it
/// does. That is the same split [`Choice`] makes.
///
/// No confirmation flag either, and that one is a finding rather than an
/// omission: a question asked *after* a control is pressed belongs to whatever
/// is holding the interaction, and a renderer that drew it would be asking
/// before there was anything to answer.
/// How a picture sits in the box it is given.
///
/// An intent rather than a value, so a renderer picks the expression it has:
/// `object-fit` in a webview, a texture's UV rect in egui, and in a terminal a
/// choice about how many cells the blit gets. Named because MNW already makes
/// the distinction deliberately at 17 sites and makes it three different ways,
/// which is a policy the app decided rather than one a shared crate would be
/// picking by accident.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Default)]
#[non_exhaustive]
pub enum Fit {
    /// The picture's own proportions, and the box takes the height they imply.
    ///
    /// The default because it is the only one that shows the whole picture at
    /// its own shape, so a renderer that ignores this enum entirely is still
    /// right about the common case. A screenshot wants this; the shipped MNW
    /// carousel sets no `object-fit` at all, which is this.
    #[default]
    Natural,
    /// Fill the box and crop whatever does not fit.
    ///
    /// For a picture in a slot whose shape the layout fixed: a thumbnail, an
    /// avatar, cover art. 15 of MNW's 17 sites.
    Cover,
    /// Fit inside the box whole, leaving space on two sides.
    ///
    /// The letterbox. For when the whole picture matters more than filling the
    /// space, and the space is not the picture's shape.
    Contain,
}

/// A picture, and what it says to someone who is not looking at it.
///
/// # No source
///
/// [`Act`]'s split, for [`Act`]'s reason. A source is an address, and this
/// crate has no notion of an address: it says what a thing *is* and the caller
/// keeps what it points at. The three findings dropped from 0.11.0 were all
/// this same shape.
///
/// It matters more here than it does for a control, because a picture is the
/// one member where the address is most of what a webview needs and *none* of
/// what the description knows. `quasi_router::Node::Image` carries the URL, the
/// way it carries an `Action` for a control.
///
/// # Why [`alt`](Self::alt) is not optional
///
/// Every other host has to draw something, and for two of the three the alt
/// text is not a fallback but the whole rendering: a terminal without a
/// graphics protocol has the words and nothing else. Making it optional would
/// make "this picture is invisible on a terminal" the default, and the
/// description would be carrying a webview assumption in its shape.
///
/// An image that genuinely says nothing — a rule, a spacer, a decoration
/// repeating what the text beside it already said — is an empty `alt`, which is
/// the same thing HTML means by it and is a claim rather than an oversight.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub struct Image<'a> {
    /// What the picture says, for anything not showing it.
    ///
    /// Empty means the picture is decorative and adds nothing to the text
    /// around it. See the type's own docs on why this is not an `Option`.
    pub alt: &'a str,
    /// A visible line under the picture, where the app wants one.
    ///
    /// Distinct from [`alt`](Self::alt) and the difference is who it is for: a
    /// caption is content everybody reads, alt text is what stands in for the
    /// picture. A screenshot captioned "The library view" still needs alt text
    /// describing what is in the shot.
    pub caption: Option<&'a str>,
    /// How it sits in the box it is given.
    pub fit: Fit,
    /// The picture's own dimensions, where the app knows them.
    ///
    /// **Not a display size**, and that distinction is what makes this belong
    /// here rather than fall foul of the deferral rule. Saying a picture should
    /// be 320 points wide is a layout value and is not the description's to
    /// give. Saying the file is 5120x3412 is a fact *about the picture*, the
    /// same kind of fact [`alt`](Self::alt) is, and no renderer can find it out
    /// without fetching the bytes.
    ///
    /// # What it is for, and it is not decoration
    ///
    /// Without it a renderer cannot reserve room, so the picture occupies
    /// nothing until it arrives and then takes its full height at once,
    /// shoving everything below it down the screen. Measured on MNW's landing
    /// page 2026-08-14: a 478px jump per frame, and a cumulative layout shift
    /// of 0.087 for the page, which is most of the way to the 0.1 that counts
    /// as bad.
    ///
    /// Every host wants it and none can derive it. A webview writes `width` and
    /// `height` so the browser holds the space; egui sizes a texture; a
    /// terminal with a graphics protocol scales a blit into cells. This was
    /// missing from 0.21.0, which is the release that added [`Image`], and its
    /// absence is the defect rather than an omission.
    ///
    /// `None` is honest and common: a creator-uploaded image whose dimensions
    /// the app never recorded genuinely does not know. It means the renderer
    /// cannot reserve, not that the picture has no size.
    pub intrinsic: Option<Extent>,
    /// Whether the picture is needed with the screen, or can arrive later.
    pub loading: Loading,
}

/// A picture's own pixel dimensions.
///
/// Deliberately not [`makeover_geometry`]'s business. Geometry answers *how
/// much space a thing should get*, which is a scale question with the same
/// answer on every screen. This is the intrinsic size of one asset, which is a
/// fact about that asset and varies per picture.
///
/// [`makeover_geometry`]: https://docs.rs/makeover-geometry
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub struct Extent {
    /// Width in the picture's own pixels.
    pub width: u32,
    /// Height in the picture's own pixels.
    pub height: u32,
}

impl Extent {
    /// A picture's dimensions.
    #[must_use]
    pub const fn new(width: u32, height: u32) -> Self {
        Self { width, height }
    }

    /// Width over height, or `None` if either side is zero.
    ///
    /// The form a renderer actually reserves space with: a box that knows its
    /// proportion holds the right height at any width, which is what a
    /// responsive picture needs and what a fixed pixel height cannot give.
    #[must_use]
    pub fn ratio(self) -> Option<f32> {
        (self.width > 0 && self.height > 0).then(|| self.width as f32 / self.height as f32)
    }
}

/// When a picture is needed.
///
/// A claim about *importance and position* rather than a fetch mechanism, which
/// is why it is the description's to make: only the app knows whether a picture
/// is the first thing on the screen or the fortieth thing down a list.
///
/// # Eager is the default, and that is a correctness choice
///
/// 0.21.0 emitted the webview's `loading="lazy"` for every picture, on the
/// evidence that the one consumer measured wrote it. That was reading a habit
/// as a rule. Deferring a picture that is on screen at first paint does not
/// save anything -- it is needed immediately either way -- and it delays the
/// arrival, so the space it eventually takes is claimed later and the shift is
/// more visible, not less.
///
/// So the safe answer is the default and the optimisation is opted into. A
/// carousel is the case that proves the two cannot be one setting for the
/// renderer to choose: its first frame is on screen and its other frames are
/// not, in the same widget, at the same moment.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Default)]
#[non_exhaustive]
pub enum Loading {
    /// Needed with the screen. Fetch it now.
    #[default]
    Eager,
    /// Not on screen yet. It can wait until it is near.
    Lazy,
}

impl<'a> Image<'a> {
    /// A picture that carries its own proportions.
    #[must_use]
    pub const fn new(alt: &'a str) -> Self {
        Self {
            alt,
            caption: None,
            fit: Fit::Natural,
            intrinsic: None,
            loading: Loading::Eager,
        }
    }

    /// The picture's own dimensions, so a renderer can hold its place.
    #[must_use]
    pub const fn intrinsic(mut self, width: u32, height: u32) -> Self {
        self.intrinsic = Some(Extent::new(width, height));
        self
    }

    /// This picture is not on screen yet; it can arrive when it is near.
    #[must_use]
    pub const fn lazy(mut self) -> Self {
        self.loading = Loading::Lazy;
        self
    }

    /// A visible line under it.
    #[must_use]
    pub const fn caption(mut self, caption: &'a str) -> Self {
        self.caption = Some(caption);
        self
    }

    /// How it sits in its box.
    #[must_use]
    pub const fn fit(mut self, fit: Fit) -> Self {
        self.fit = fit;
        self
    }

    /// Whether the picture adds anything for someone not looking at it.
    ///
    /// A renderer with no way to show a picture uses this to decide between
    /// drawing the alt text and drawing nothing at all. Both are correct and
    /// the difference is this flag: standing in for a decorative rule with the
    /// word "decoration" is worse than leaving the space empty.
    #[must_use]
    pub const fn speaks(self) -> bool {
        !self.alt.is_empty()
    }
}

#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub struct Act<'a> {
    /// What the control says.
    pub label: &'a str,
    /// The key that reaches it where a host has keys.
    ///
    /// The one member written for a terminal before there was one. A webview
    /// hangs it off `accesskey` or ignores it; a terminal has nothing else to
    /// offer, so this is the whole of how a control is reached there.
    pub key: Option<&'a str>,
    /// What pressing it means. [`Tone::Danger`] is the destructive one.
    pub tone: Tone,
    /// Disabled, or nothing said.
    ///
    /// [`State::Disabled`] is what changes what a renderer may do: see
    /// [`State::suppresses_interaction`], which is what says a disabled control
    /// is drawn and not reachable. It has been the only member since 0.19.0,
    /// and a control's focus is not sayable here at all — see the crate header,
    /// "Reach, focus and the focus ring".
    pub state: Option<State>,
}

impl<'a> Act<'a> {
    /// An ordinary control, reachable, with no key.
    #[must_use]
    pub const fn new(label: &'a str) -> Self {
        Self {
            label,
            key: None,
            tone: Tone::Neutral,
            state: None,
        }
    }

    /// The key that reaches it.
    #[must_use]
    pub const fn key(mut self, key: &'a str) -> Self {
        self.key = Some(key);
        self
    }

    /// What pressing it means.
    #[must_use]
    pub const fn tone(mut self, tone: Tone) -> Self {
        self.tone = tone;
        self
    }

    /// Focus, or disabled.
    #[must_use]
    pub const fn state(mut self, state: State) -> Self {
        self.state = Some(state);
        self
    }

    /// Whether the control is drawn and does not answer.
    #[must_use]
    pub fn disabled(&self) -> bool {
        self.state.is_some_and(State::suppresses_interaction)
    }
}

/// A named part of a screen.
///
/// The thing `makeover-geometry` deliberately does not name: it names the space
/// *between* things by relationship, and nothing named the things. Six named
/// members, taken from what the two webview apps actually use, plus
/// [`Region::Bespoke`] for the parts no description should reach. Both apps'
/// `layout.css` currently names exactly two things, `.raised` and `.well`, so
/// this layer is absent rather than divergent, which makes it the cheapest of
/// the schemas to add and the easiest to over-build.
///
/// `#[non_exhaustive]` arrives with [`Region::Widget`], the pairing [`RowPart`]
/// made at 0.9.0 and [`Readiness`] at 0.12.0, and for the same reason: the
/// member after this one should not be a lockstep event across three renderers.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub enum Region<'a> {
    /// A full-width strip with a title slot and an actions cluster, either of
    /// which may be empty. goingson's `.page-header`, Balanced Breakfast's
    /// `.header` and `.detail-header` are all this, differing only in which
    /// slots they fill.
    Band,
    /// A persistent column beside the content, holding navigation.
    Sidebar,
    /// A region of content with its own scroll.
    Pane,
    /// Two panes side by side, where the left chooses what the right shows.
    Split,
    /// A set of panes, one visible at a time, with a [`Selector::Tabs`] above.
    TabGroup,
    /// Content over a scrim, taking input until dismissed.
    Modal,
    /// A region this crate names the *place* of and nothing else. The app owns
    /// what goes in it.
    ///
    /// The escape hatch, and the thing that keeps the description honest about
    /// its own limits. A day-plan timeline, a kanban board, a calendar and the
    /// paint interaction over the timeline are not describable here and are not
    /// going to become describable: a description expressive enough to produce
    /// a timeline is a widget library wearing a description's name.
    ///
    /// But a screen containing one still has to be a screen. Without this
    /// member the description covers only the boring screens, and the four that
    /// make goingson worth using would need a second, undescribed path beside
    /// the router. Two paths is how the vocabulary starts drifting from the app
    /// again, which is the exact failure this crate exists to end.
    ///
    /// So the description says "a thing called `day-plan` goes here" and stops.
    /// The name is opaque: this crate never interprets it, and no renderer is
    /// expected to know what it means beyond handing the space over.
    Bespoke {
        /// What the app calls it. Never interpreted here.
        name: &'a str,
    },
    /// A named assembly of things the vocabulary already says.
    ///
    /// The third tier, between a primitive and [`Bespoke`](Self::Bespoke).
    /// Stated by Max 2026-08-12 answering the carousel: "something in between a
    /// primitive and a bespoke interface, like a widget, which is just an
    /// assembly of primitives." Full note: wiki `widget-tier`.
    ///
    /// # What separates it from the two members either side
    ///
    /// A primitive is a thing every renderer draws from scratch, and the test
    /// it has to pass is that every host has an honest answer. A carousel fails
    /// that test — a terminal has no carousel — which is the same refusal
    /// `Node::Html` got and is why the carousel sat unsayable for months.
    ///
    /// [`Bespoke`](Self::Bespoke) fails it from the other side. Bespoke is for
    /// what one app owns and nobody will build twice, and it carries *no*
    /// contents: the description names the place and stops. A carousel is
    /// furniture any app would have, and every part of it — an ordered set of
    /// frames, a position, prev and next, a strip of position indicators — is
    /// already sayable. Only the assembly had no name.
    ///
    /// So this member is the pair the other two are not: a name **and**
    /// contents. The contents are the assembly, in the region's own body, said
    /// in members that already exist.
    ///
    /// # Why the name does not have to be understood
    ///
    /// A renderer that recognises the name draws it the way its host does it: a
    /// carousel in a webview, a pager with a count in a terminal, a selector in
    /// egui. A renderer that does not recognise it walks the body, which is
    /// primitives all the way down and which it can already draw.
    ///
    /// That is what lets the widget set be **open** without every renderer
    /// knowing every widget. An unrecognised widget degrades to its assembly
    /// instead of failing, so a second or third party can name one without
    /// three renderers releasing in lockstep to accept it. Contrast
    /// [`Bespoke`](Self::Bespoke), which no renderer can degrade: there is
    /// nothing under it to fall back to.
    ///
    /// # What it does not do
    ///
    /// It does not make a timeline describable, and the refusal in the crate
    /// header stands unchanged. A widget is an assembly of things the
    /// vocabulary *already* says; anything that needs a member the vocabulary
    /// does not have is a finding about the vocabulary or it is
    /// [`Bespoke`](Self::Bespoke). A widget is never the way a primitive gets
    /// added by the back door.
    Widget {
        /// What the assembly is called. This crate never interprets it, and a
        /// renderer is free not to know it.
        name: &'a str,
    },
}

impl<'a> Region<'a> {
    /// How the region sits on what is behind it.
    #[must_use]
    pub const fn depth(self) -> Depth {
        match self {
            Self::Band | Self::Sidebar | Self::Split | Self::TabGroup => Depth::Flat,
            // A pane is looked into, the same as a table body or a tag tree.
            Self::Pane => Depth::Well,
            Self::Modal => Depth::Raised,
            // Flat because it inherits: a bespoke region takes the depth of
            // whatever frames it. An app that wants its timeline in a well puts
            // it in a `Pane`, which composes rather than adding a knob here.
            //
            // A widget inherits for the same reason and it matters more here,
            // because a widget is drawn by whichever renderer recognises it. A
            // depth set here would be this crate deciding that a carousel is
            // raised on every host, which is the kind of value the deferral
            // rule exists to refuse.
            Self::Bespoke { .. } | Self::Widget { .. } => Depth::Flat,
        }
    }

    /// Whether this crate can say anything about the region's contents.
    ///
    /// A renderer walks the description and hands every region it understands
    /// to the right drawing code. This is how it tells the two apart, and the
    /// reason it is a method rather than a `matches!` at each renderer: there
    /// is exactly one opaque member and there should stay exactly one.
    ///
    /// [`Widget`](Self::Widget) is described, and that is the whole of what
    /// separates it from [`Bespoke`](Self::Bespoke) here. Both carry a name
    /// this crate never interprets; only one of them carries contents under it.
    /// A renderer that does not recognise a widget's name still walks its body,
    /// so there is nothing for it to hand over and nothing it cannot draw.
    #[must_use]
    pub const fn described(self) -> bool {
        !matches!(self, Self::Bespoke { .. })
    }

    /// The name an app gave this region, if it gave one.
    ///
    /// [`Bespoke`](Self::Bespoke) and [`Widget`](Self::Widget) are the two
    /// members that carry a name, for two different purposes: one says what the
    /// app will fill the space with, the other says what the assembly under it
    /// is called. A renderer dispatching on either wants the string without
    /// caring which member it came from, and writing that `matches!` at each
    /// renderer is how the two drift apart.
    #[must_use]
    pub const fn name(self) -> Option<&'a str> {
        match self {
            Self::Bespoke { name } | Self::Widget { name } => Some(name),
            // Spelled out rather than a wildcard, so a member added later has
            // to answer whether it carries a name instead of inheriting `None`
            // by sitting under a `_`.
            Self::Band
            | Self::Sidebar
            | Self::Pane
            | Self::Split
            | Self::TabGroup
            | Self::Modal => None,
        }
    }
}

/// How many of a region's children are visible at once.
///
/// `4dcd241b`. Three findings turned out to be one sentence the vocabulary
/// could not say: *this region holds several children and shows some of them,
/// and the reader can change which.* [`Region::TabGroup`] existed with nothing
/// saying which tab was open, a carousel had nothing saying which frame was up,
/// and a disclosure had nothing saying whether its one child was showing at all.
///
/// Because the fact lived nowhere, a renderer had two moves: hardcode a widget
/// name, or draw every child. That is what put per-widget code in renderers, and
/// it was the missing member rather than the widget tier that put it there.
///
/// # What is here and what is not
///
/// The *kind*, and only the kind. Which child is currently up is the current
/// answer, and a layer that defers every address does not hold the current
/// answer either — the split [`Selector`] already makes, where this crate says
/// what kind of chooser a thing is and the router says which option is picked.
/// So a holder of regions carries the index and the per-child label beside this.
///
/// # What a renderer does with it
///
/// Derives its chrome, once, for every widget rather than per name:
///
/// - Children carrying labels get a strip of the labels, the current one marked.
/// - Children carrying none get previous, position, next.
/// - [`AtMostOne`](Self::AtMostOne) over one child gets a summary line that
///   opens.
///
/// The name on [`Region::Widget`] survives as app vocabulary, for a renderer
/// that wants to do something *special* with one, which is what it should have
/// been from the start.
///
/// Degradation runs the way it already did: a renderer ignoring this draws every
/// child, which is more content rather than less.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Default)]
#[non_exhaustive]
pub enum Showing {
    /// Every child, in order. What every region did before this existed.
    #[default]
    All,
    /// Exactly one. A carousel, a tab group.
    One,
    /// One, or none. A disclosure, which is closed until it is opened.
    AtMostOne,
}

impl Showing {
    /// Whether the reader can change which child is up.
    ///
    /// The question every renderer's region arm asks before deriving any
    /// chrome, and a method rather than a `matches!` at each renderer for
    /// [`Region::name`]'s reason: three renderers writing the same comparison is
    /// how they come to disagree about a member added later.
    #[must_use]
    pub const fn selective(self) -> bool {
        !matches!(self, Self::All)
    }

    /// Whether showing nothing is a legal state.
    ///
    /// True only for [`AtMostOne`](Self::AtMostOne). A renderer needs this to
    /// know whether its control closes as well as moves: a carousel's row moves
    /// between frames and never reaches empty, and a disclosure's summary line
    /// is the same control wearing its closed state.
    #[must_use]
    pub const fn dismissible(self) -> bool {
        matches!(self, Self::AtMostOne)
    }
}

/// How much of the width an arrangement's first region takes.
///
/// `e0fd485e`. Nothing said how much room a region got, so every renderer
/// invented its own number and two hosts showing one screen disagreed about
/// its proportions. A webview never noticed, because the stylesheet answered
/// once for every consumer; a terminal has no stylesheet to inherit from, so
/// `quasi-tui` picked 24 columns for a sidebar and 40% for a list pane and
/// neither had anything behind it.
///
/// # A proportion, never a unit
///
/// Held as a percentage, and that is the only form it comes in. A description
/// carrying columns would be describing a terminal and one carrying pixels a
/// webview, and the whole point is that both honour the same fact: a terminal
/// resolves it against a column count, a webview writes it into a grid, and
/// neither has to know what the other did.
///
/// It is not [`makeover_geometry::Ratio`]'s job either, which was the first
/// guess. Geometry is scales that answer the same for every screen and takes
/// no input that would let a sidebar screen differ from a list-detail one.
///
/// [`makeover_geometry::Ratio`]: https://docs.rs/makeover-geometry
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, PartialOrd, Ord)]
pub struct Share(u8);

impl Share {
    /// What a sidebar takes, when nobody says otherwise.
    ///
    /// A quarter. `quasi-tui` drew 24 columns, which is a quarter of a
    /// 96-column terminal and about a fifth of a wide one; a quarter is that
    /// number said in the form a webview can honour too.
    pub const SIDEBAR: Self = Self(25);

    /// What the list side of a list-detail takes, when nobody says otherwise.
    ///
    /// `quasi-tui`'s 40%, which was already a proportion and is the one number
    /// this member did not have to invent.
    pub const LIST: Self = Self(40);

    /// A share of the width, as a percentage.
    ///
    /// Clamped to 5..=95 rather than refused. A description that asked for a
    /// region of nothing is a bug in the app, and a renderer drawing a region
    /// zero cells wide reports it as a region that vanished, which is the
    /// hardest kind of bug to find from what is on the screen.
    #[must_use]
    pub const fn percent(percent: u8) -> Self {
        Self(if percent < 5 {
            5
        } else if percent > 95 {
            95
        } else {
            percent
        })
    }

    /// The share as a percentage.
    #[must_use]
    pub const fn as_percent(self) -> u8 {
        self.0
    }

    /// This share of a width, rounded to the nearest whole unit.
    ///
    /// What a terminal calls to turn the proportion into columns. At least one,
    /// because a region the description named should be visible: a screen
    /// 3 columns wide is unusable either way, and a sidebar that is there is a
    /// truer picture of the description than a sidebar that is not.
    #[must_use]
    pub const fn of(self, whole: u16) -> u16 {
        let taken = (whole as u32 * self.0 as u32).div_ceil(100);
        if taken == 0 { 1 } else { taken as u16 }
    }
}

/// How a screen is laid out.
///
/// Two, and the second is not a variant of the first. goingson is list-detail,
/// Balanced Breakfast is sidebar plus content, and neither app has a third.
/// The tab group is a modifier rather than a member, because goingson uses it
/// *inside* the same content region rather than instead of one.
///
/// This exists at all because the router has to be able to express a screen
/// rather than only a control. Discovering the arrangement layer missing after
/// the renderers exist is a redesign; naming two now is a morning.
///
/// # Why the share rides here
///
/// `e0fd485e`. A share is per-arrangement: how much a sidebar takes and how
/// much a list side takes are different questions, and this enum is the only
/// thing that knows which one is being asked. Geometry would have had to invent
/// a channel to be told.
///
/// [`list_detail`](Self::list_detail) and
/// [`sidebar_content`](Self::sidebar_content) build these with the default
/// shares, so a screen that has no opinion does not have to have one.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum Arrangement {
    /// A list that chooses what the detail beside it shows.
    ListDetail {
        /// Whether the detail side is a [`Region::TabGroup`].
        tabbed: bool,
        /// How much of the width the list side takes.
        share: Share,
    },
    /// Navigation down the side, content filling the rest.
    SidebarContent {
        /// How much of the width the sidebar takes.
        share: Share,
    },
}

impl Arrangement {
    /// A list and a detail beside it, at the default share.
    #[must_use]
    pub const fn list_detail(tabbed: bool) -> Self {
        Self::ListDetail {
            tabbed,
            share: Share::LIST,
        }
    }

    /// A sidebar and content beside it, at the default share.
    #[must_use]
    pub const fn sidebar_content() -> Self {
        Self::SidebarContent {
            share: Share::SIDEBAR,
        }
    }

    /// How much of the width the first region takes.
    #[must_use]
    pub const fn share(self) -> Share {
        match self {
            Self::ListDetail { share, .. } | Self::SidebarContent { share } => share,
        }
    }

    /// The same arrangement, at this share.
    #[must_use]
    pub const fn with_share(self, share: Share) -> Self {
        match self {
            Self::ListDetail { tabbed, .. } => Self::ListDetail { tabbed, share },
            Self::SidebarContent { .. } => Self::SidebarContent { share },
        }
    }
}

/// How wide the content of a whole screen runs.
///
/// `0eccff0d`, and [`Share`]'s sibling one level up: that one says how a
/// screen's width is divided between regions, this says how much of the window
/// the screen uses in the first place. Both are the description's, which is
/// what answering the two together settled.
///
/// Measured in the MNW server, where 69 of 72 templates carry exactly one of
/// three mutually exclusive classes and the choice is per screen. GoingsOn
/// reaches for `max-width` 56 times and Balanced Breakfast 12, neither with a
/// token for it, so three apps were solving one thing by hand.
///
/// # Named for the measure, not for MNW's classes
///
/// A renderer that is not a browser has to answer this too, and `padded-page`
/// tells a terminal nothing. The three say how wide the text runs, which is a
/// question every renderer can answer: a webview with a `max-width`, a terminal
/// with gutters, an immediate-mode frame with its own width.
///
/// `#[non_exhaustive]` for [`Fill`]'s reason. The set is closed today because
/// the measurement found three, and a fourth arriving should not be a lockstep
/// release across nine repos.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Default)]
#[non_exhaustive]
pub enum Measure {
    /// The whole width, with gutters. The default, and 53 of the 69.
    ///
    /// What a dashboard, a table and a settings screen want: the content is
    /// wide because the content *is* wide, and constraining it would waste the
    /// window.
    #[default]
    Wide,
    /// Capped at a comfortable page width, centred. 13 of the 69.
    ///
    /// A form, a sign-in, a purchase. Content that does not get better by
    /// getting wider, but is not prose either.
    Contained,
    /// Capped at a line length that reads well. 3 of the 69.
    ///
    /// Prose. The narrowest of the three, and the one with a reason outside
    /// taste: a line of text past roughly 75 characters costs the reader the
    /// return sweep.
    Reading,
}

impl Measure {
    /// A stable name, for a renderer that needs to spell it.
    ///
    /// Here rather than in each renderer for [`Sort::as_str`]'s reason: three
    /// renderers spelling one enum is three chances to spell it differently.
    #[must_use]
    pub const fn as_str(self) -> &'static str {
        match self {
            Self::Wide => "wide",
            Self::Contained => "contained",
            Self::Reading => "reading",
        }
    }
}

/// What kind of value a form field takes.
///
/// The union of the two vocabularies that diverged, which is what triggered
/// this crate. They have since converged on their own: both apps now have a
/// `renderFormField` emitting the same anatomy, and what is left differing is
/// the kind set, the error shape, and whether the return is a string or a node.
///
/// Validation is deliberately absent. Neither app has a shared story (goingson
/// validates after collecting the form data, with per-field transform hooks;
/// Balanced Breakfast has `required` and nothing else), and a schema that
/// describes fields but not constraints acquires a constraint layer per app,
/// which is exactly how the current divergence started. Naming it absent is a
/// decision; leaving it unmentioned would not be.
/// `#[non_exhaustive]` for the reason [`Fill`] is: renderers match on this and
/// the set keeps growing, so growth must not be a lockstep event. Email, Url
/// and Tel arriving in 0.5.0 is the second growth in two releases.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub enum FieldKind {
    /// A single line of text.
    Text,
    /// A single line of text that must never be echoed, logged or round-tripped
    /// through anything that might persist it.
    Secret,
    /// A number.
    Number,
    /// An email address.
    ///
    /// Distinct from [`Text`](Self::Text) because the distinction is not
    /// decoration: a webview renderer emits `type="email"`, which on a touch
    /// device changes the keyboard that appears and turns on the platform's own
    /// validation. goingson ships to iOS, so collapsing this into text costs a
    /// keyboard with no `@` on it.
    ///
    /// Added 0.5.0, from goingson's contact form.
    Email,
    /// A URL. Same reasoning as [`Email`](Self::Email).
    ///
    /// Added 0.5.0, from goingson's contact-social and contact-feed forms.
    Url,
    /// A telephone number. Same reasoning as [`Email`](Self::Email), and the
    /// clearest case of it: the keyboard is a numeric pad rather than letters.
    ///
    /// Added 0.5.0, from goingson's contact-phone form.
    Tel,
    /// A calendar day, with no time of day in it.
    ///
    /// [`Email`](Self::Email)'s argument, and it carries further: a webview
    /// emits `type="date"`, which is a native picker, the platform's own
    /// validation, and on a touch device the date keyboard. Described as
    /// [`Text`](Self::Text) with a hint reading "YYYY-MM-DD", all three are
    /// lost and the hint is doing the platform's job in prose.
    ///
    /// The membership test passes on every host without stretching: a webview
    /// and a Tauri app emit the input, egui has a date picker, a terminal
    /// prompts for a day and can validate it, a CLI takes an argument.
    ///
    /// # The value is ISO 8601, `YYYY-MM-DD`
    ///
    /// Named here rather than left to each host, because a host that picks
    /// differently sends a server something it parses differently, and the
    /// failure is silent and per-host. It is `<input type="date">`'s own wire
    /// format, so the webview renderer owes nothing to honour it and the other
    /// hosts have one spelling to meet. [`DATE_FORMAT`] is the constant, and a
    /// test asserts this doc and that constant agree.
    ///
    /// Added 0.15.0, from the MNW server's git access-token expiry
    /// (`user_ssh_keys_tab.html`) and six further sites across the server and
    /// goingson.
    Date,
    /// A calendar day and a time of day together.
    ///
    /// Apart from [`Date`](Self::Date) because the question is different rather
    /// than more precise: "which day does this expire" and "at what moment does
    /// this publish" are asked by different screens and answered by different
    /// controls. A webview emits `type="datetime-local"` for one and
    /// `type="date"` for the other, and a host that collapsed them would ask
    /// half the tree for a precision it does not want.
    ///
    /// Both arrived together on measurement rather than on symmetry: 13 sites
    /// of each across the MNW server and goingson, and **zero** of `time`,
    /// `month` or `week`, which is why those are not here. A member added for a
    /// case nobody has is a member designed against nothing, which is
    /// [`File`](Self::File)'s reasoning about `accept` applied to a whole
    /// member.
    ///
    /// # The value is `YYYY-MM-DDTHH:MM`, local, with no zone
    ///
    /// `<input type="datetime-local">`'s own format, and the "local" is the
    /// load-bearing half: the value carries no offset and no `Z`, so the moment
    /// it names is only fixed once something supplies a zone. That is the app's
    /// business and not the description's. Seconds are absent, which is the
    /// browser's own default and is left as the rule rather than restated as a
    /// constraint. [`DATETIME_FORMAT`] is the constant.
    ///
    /// [`Field::min`] and [`Field::max`] already take "the host's own spelling
    /// of a bound", so a floor of *not in the past* needs nothing new here: it
    /// is a string in this same format.
    ///
    /// Added 0.15.0, from goingson's snooze picker and day planner and the MNW
    /// server's publish-at fields.
    DateTime,
    /// Several lines of text.
    Textarea,
    /// One of a fixed set, offered behind a control that shows one at a time.
    Select,
    /// One of a fixed set, with every option on screen at once.
    ///
    /// Not a presentation of [`Select`](Self::Select), which is the reading to
    /// resist: what differs is a property of the *question*. A choice that is
    /// consequential or irreversible has to be readable without opening
    /// anything, because a closed control shows one option and hides the rest,
    /// and the one it shows is whichever was current before the user had read
    /// the alternatives. audiofiles asks whether a library copies samples into
    /// its store or references them where they lie — which cannot be changed
    /// afterwards — and had already promoted that out of a checkbox by hand,
    /// with a comment giving this reason, before the description could say it.
    ///
    /// It was described here at 0.8.1 as "the one HTML input type this enum was
    /// missing", which was not true then and is not true now: `file` arrived at
    /// 0.11.0 and `date` and `datetime-local` at 0.15.0. Everything here is
    /// still an `<input type=...>`, a `<select>` or a `<textarea>`, and the way
    /// this enum grows is by a site being measured rather than by a list being
    /// completed, so "the last one" is not a claim it should make again.
    ///
    /// Added 0.8.1, from audiofiles' Add Library form.
    Radio,
    /// On or off.
    Checkbox,
    /// A file the user picks from wherever the host keeps files.
    ///
    /// Added 0.11.0, `844b5ae0`, from goingson's project-dashboard attachments
    /// column. It was filed as a router finding — a control whose destination is
    /// a host capability rather than an address — and splitting it is what made
    /// it two answers instead of one member satisfying neither. *Opening* a file
    /// is a one-way handoff and needs no new API. *Picking* one returns a value
    /// into a write, which is a form concern, which is this.
    ///
    /// The membership test passes on every host and not by a stretch: a Tauri
    /// app opens a native picker, a server renders `<input type="file">`, a
    /// terminal prompts for a path, a CLI takes an argument. That is closer to
    /// [`Email`](Self::Email), which exists because it changes the keyboard,
    /// than to anything bespoke.
    ///
    /// It carries no accepted-types list and no multiple flag, and that is
    /// measured rather than deferred: `accept` appears at zero sites in either
    /// app. A member added for a case nobody has is a member designed against
    /// nothing.
    File,
    /// Carried through the form and never shown.
    Hidden,
}

/// The wire format a [`FieldKind::Date`] value takes: ISO 8601, `YYYY-MM-DD`.
///
/// A constant rather than a sentence in a doc comment, because the reason to
/// name the format at all is that a host picking its own would fail silently
/// against a server parsing another. A host that cannot emit the native control
/// still has one spelling to meet, and can say which one it meant.
pub const DATE_FORMAT: &str = "%Y-%m-%d";

/// The wire format a [`FieldKind::DateTime`] value takes: `YYYY-MM-DDTHH:MM`,
/// local, carrying no zone and no seconds.
///
/// [`DATE_FORMAT`]'s sibling and there for its reason. The absent zone is a
/// property of the value rather than an omission: the moment is not fixed until
/// something outside the description supplies one.
pub const DATETIME_FORMAT: &str = "%Y-%m-%dT%H:%M";

impl FieldKind {
    /// Whether the value the kind takes is a moment rather than a string.
    ///
    /// Named once here for the reason [`offers_options`](Self::offers_options)
    /// is: two kinds answer yes, and a host that has to parse or format a value
    /// needs to ask without spelling the pair out at each renderer. A third
    /// temporal kind should land here and nowhere else.
    ///
    /// The format each one takes is [`DATE_FORMAT`] and [`DATETIME_FORMAT`].
    #[must_use]
    pub const fn temporal(self) -> bool {
        matches!(self, Self::Date | Self::DateTime)
    }

    /// Whether the field is drawn at all.
    #[must_use]
    pub const fn visible(self) -> bool {
        !matches!(self, Self::Hidden)
    }

    /// Whether the value must be kept out of logs and diagnostics.
    #[must_use]
    pub const fn confidential(self) -> bool {
        matches!(self, Self::Secret)
    }

    /// Where the field's own label sits.
    ///
    /// A checkbox labels itself on the right of the box; everything else takes
    /// a label above. Both webview apps already do this and both special-case
    /// it inline, which is the tell that it belongs in the description.
    ///
    /// A [`Radio`](Self::Radio) is not one of them, and the near-miss is worth
    /// naming: its *options* each label themselves, but the field still asks a
    /// question above them, so the group takes a label like everything else.
    #[must_use]
    pub const fn labels_itself(self) -> bool {
        matches!(self, Self::Checkbox)
    }

    /// Whether the kind reads [`Field::options`].
    ///
    /// Two kinds do, so the pair is named once here rather than spelled out at
    /// each renderer and again in [`Field::options`]' own doc, where "every
    /// kind but `Select`" was true for exactly one release. A third
    /// option-taking kind should land here and nowhere else.
    #[must_use]
    pub const fn offers_options(self) -> bool {
        matches!(self, Self::Select | Self::Radio)
    }
}

/// One option offered by a field [`FieldKind::offers_options`] accepts.
///
/// Two strings, because the submitted value and the read label are different
/// facts and every renderer that has tried to collapse them has had to
/// un-collapse them later. `makeover-webview` invented this shape writing its
/// form emitter and it is taken here unchanged; moving it down rather than
/// re-deriving it is the point, since the second and third renderers were each
/// going to arrive at a near-miss of it.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub struct Choice<'a> {
    /// What is submitted.
    pub value: &'a str,
    /// What is read.
    pub label: &'a str,
}

impl<'a> Choice<'a> {
    /// An option whose submitted value is also its label.
    #[must_use]
    pub const fn plain(value: &'a str) -> Self {
        Self {
            value,
            label: value,
        }
    }
}

/// One field of a form.
///
/// Borrowed rather than owned: a description is built, read once by a renderer,
/// and dropped. Nothing here outlives the screen it describes.
///
/// # What it carries, and what it does not
///
/// Stated here so the next renderer does not re-ask, which is what the first
/// two both did. It carries everything a renderer needs to *draw* the field:
/// its kind, what it is called, what it is asked for, its standing help, what
/// is wrong with it now, whether it is compulsory, whether it hides behind a
/// disclosure, its ghost text, and the options it offers.
///
/// It does not carry the **current value**, and it is not going to. That is the
/// one thing here that is genuinely renderer state: a webview reads it back out
/// of the DOM, an immediate-mode renderer holds a `&mut` to the app's own field
/// and writes through it, and a terminal keeps an edit buffer. A description
/// that carried the value would have to carry a way to write it back, at which
/// point it is a form model and no longer a description.
///
/// **Constraints** are here and enforcement is not, which is one line rather
/// than two. [`required`], [`max_length`], [`min`] and [`max`] are facts about
/// the *question*, so a renderer can emit its host's idiom for each — an HTML
/// attribute, a marked label, a clamped spinner — and the platform helps the
/// user before anything is submitted. Deciding that a value is wrong stays with
/// whoever validated, and [`error`] is that decision arriving back.
///
/// The set stops before `pattern`, and stops there on both tests at once. A
/// regex has an honest answer in a webview and none anywhere else: egui would
/// have to run it per keystroke and decide what a half-typed value means, which
/// is enforcement wearing description's clothes. And it is one site in goingson
/// and none in Balanced Breakfast, against 8 and 1 for `maxlength`. Measured
/// 2026-08-09, `2cbad3e2`.
///
/// [`error`]: Field::error
/// [`required`]: Field::required
/// [`max_length`]: Field::max_length
/// [`min`]: Field::min
/// [`max`]: Field::max
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub struct Field<'a> {
    /// What kind of value it takes.
    pub kind: FieldKind,
    /// The name the value is submitted under.
    pub name: &'a str,
    /// What the user is asked for.
    pub label: &'a str,
    /// Standing help, shown whether or not anything is wrong.
    pub hint: Option<&'a str>,
    /// What is currently wrong with the value.
    pub error: Option<&'a str>,
    /// Ghost text shown while the field is empty.
    ///
    /// User-facing text, and it sits with `label` and `hint` rather than with
    /// the value because it is a property of the *question* and not of the
    /// answer. It lived renderer-side in `makeover-webview` until 0.8.0 for one
    /// reason and it was not a reading on where it belonged: adding a field to
    /// a published struct is a breaking change.
    ///
    /// Not a substitute for a label. A field labelled only by its placeholder
    /// loses its label the moment anything is typed, and no renderer here can
    /// make that not happen, so the description keeps both.
    pub placeholder: Option<&'a str>,
    /// The options offered, in the order they are offered.
    ///
    /// Empty for every kind [`FieldKind::offers_options`] rejects. A field
    /// described with no options is sayable on purpose: it is what an app with
    /// an unfinished-loading option list actually has, and a renderer showing
    /// an empty control says so on screen rather than in a log.
    ///
    /// Which option is *current* is not here. That is the value, and the value
    /// is renderer state.
    pub options: &'a [Choice<'a>],
    /// Whether the form refuses to submit without it.
    pub required: bool,
    /// The longest the value may be, in characters.
    ///
    /// Added 0.11.0 with [`min`](Self::min) and [`max`](Self::max), joining
    /// [`required`](Self::required), which had been the only constraint here
    /// since before the crate wrote down that it carried none.
    pub max_length: Option<u32>,
    /// The lowest value accepted, as the host would write it.
    ///
    /// Text rather than a number, because the bound is only a number for some
    /// of the kinds that take one. goingson's own sites are `min="1"` on a
    /// duration and `min="2026-08-09T14:30"` on a datetime, and a numeric member
    /// could say the first and not the second. The [`kind`](Self::kind) already
    /// says how to read it, the same way it does for the value.
    pub min: Option<&'a str>,
    /// The highest value accepted, as the host would write it. See
    /// [`min`](Self::min).
    pub max: Option<&'a str>,
    /// Whether the field lives behind a "more options" disclosure.
    pub extended: bool,
}

impl<'a> Field<'a> {
    /// A plain required-nothing field of the given kind.
    #[must_use]
    pub const fn new(kind: FieldKind, name: &'a str, label: &'a str) -> Self {
        Self {
            kind,
            name,
            label,
            hint: None,
            error: None,
            placeholder: None,
            options: &[],
            required: false,
            max_length: None,
            min: None,
            max: None,
            extended: false,
        }
    }

    /// A select offering the given options.
    ///
    /// One of the two kinds under-described by [`Field::new`], so it gets a
    /// constructor rather than leaving every call site to remember that a
    /// select with an empty `options` renders as an empty select.
    #[must_use]
    pub const fn select(name: &'a str, label: &'a str, options: &'a [Choice<'a>]) -> Self {
        Self::offering(FieldKind::Select, name, label, options)
    }

    /// A radio group offering the given options.
    ///
    /// The other. Same hazard as [`select`](Self::select) and a worse one: a
    /// radio group with no options draws nothing at all, so a call site that
    /// forgot them has an empty rectangle rather than a visibly empty control.
    #[must_use]
    pub const fn radio(name: &'a str, label: &'a str, options: &'a [Choice<'a>]) -> Self {
        Self::offering(FieldKind::Radio, name, label, options)
    }

    /// The shared body of the two constructors that take options.
    ///
    /// Private, and keyed on the kind rather than exposed, because the two
    /// public names are the point: a call site says which question it is
    /// asking, not which flag it is setting.
    const fn offering(
        kind: FieldKind,
        name: &'a str,
        label: &'a str,
        options: &'a [Choice<'a>],
    ) -> Self {
        Self {
            options,
            ..Self::new(kind, name, label)
        }
    }

    /// Whether the field is currently reporting a problem.
    ///
    /// Read this rather than testing `error.is_some()` at each renderer: the
    /// error state has to mark the field's whole group and not only the
    /// message, because a renderer with no descendant selectors (egui, a
    /// terminal) cannot find the group from the message. goingson already marks
    /// the group and Balanced Breakfast does not, so goingson's shape is the
    /// one taken here.
    #[must_use]
    pub const fn invalid(&self) -> bool {
        self.error.is_some()
    }
}

/// How much room a column asks for.
///
/// An intent, so the actual floor stays with `makeover-geometry`. goingson's
/// task table spells these as `minmax(200px, 1fr)`, `140px` and content-sized;
/// only the first three words of that survive deferral.
/// `#[non_exhaustive]`, for the reason [`Fill`] and [`FieldKind`] are: a
/// renderer matches on this and a vocabulary that grows must not break every
/// renderer when it does.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub enum Width {
    /// Takes what it needs and no more.
    Content,
    /// A fixed share, the same at every width.
    Fixed,
    /// Absorbs whatever is left over.
    Fill,
}

/// What a column is worth when there is not room for all of them.
///
/// Ordered: [`Priority::Optional`] drops first, [`Priority::Essential`] never
/// drops. This replaces addressing columns by position, which is what both
/// webview apps do today and is a live bug rather than only verbosity. goingson
/// hides mobile columns with `nth-child(n+5)` against a seven-column table, so
/// inserting a column silently hides the wrong one.
/// `#[non_exhaustive]`, same reasoning as [`Width`]. Note the ordering is the
/// whole point of the type, so a new tier has to be declared in its place in
/// the sequence rather than appended.
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
#[non_exhaustive]
pub enum Priority {
    /// Dropped first.
    Optional,
    /// Dropped once the optional columns are gone.
    Secondary,
    /// Never dropped. Without it the row does not identify itself.
    Essential,
}

/// One column of a table.
///
/// Described once. The grid track, the cell order and the drop behaviour are
/// all derived from this, rather than being three hand-written encodings that
/// must agree and are never checked against each other.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub struct Column<'a> {
    /// The heading, and the name the cell is addressed by.
    pub name: &'a str,
    /// How much room it asks for.
    pub width: Width,
    /// What it is worth when room runs out.
    pub priority: Priority,
    /// Whether the user can reorder the table by this column.
    ///
    /// `ce620871`. What reordering *calls* is not here — that is an address, and
    /// this crate names none — so a host pairs this with the route the way it
    /// pairs a row's parts with the row's activation. This says the affordance
    /// exists, which is what a renderer needs to draw a header a user can press
    /// rather than a heading they cannot.
    pub sortable: bool,
    /// Which way the table is ordered by this column, if it is.
    ///
    /// `None` on every column but the one in force. A renderer draws the caret
    /// from this and a webview sets `aria-sort`, which is why it is per column
    /// rather than a single fact on the table: the host idiom is a property of
    /// the header cell.
    ///
    /// Independent of [`sortable`](Self::sortable) rather than implied by it,
    /// because both combinations mean something. A column sorted and not
    /// sortable is a list ordered by a key the user cannot change, which is a
    /// real thing to describe and a caret worth drawing.
    pub sorted: Option<Sort>,
}

/// Which way a column is ordered.
///
/// Two, because there is no third. "Unsorted" is [`Column::sorted`] being
/// `None`, and folding it in here would be the same absence said twice.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum Sort {
    /// Smallest, earliest or first alphabetically at the top.
    Ascending,
    /// The other way.
    Descending,
}

impl Sort {
    /// The other direction, for a header that flips when pressed.
    #[must_use]
    pub const fn reversed(self) -> Self {
        match self {
            Self::Ascending => Self::Descending,
            Self::Descending => Self::Ascending,
        }
    }

    /// What a webview writes into `aria-sort`.
    ///
    /// Named here rather than in the webview renderer because a terminal and an
    /// immediate-mode painter both want the same two words for a caret's label,
    /// and three renderers picking their own is the drift this crate ends.
    #[must_use]
    pub const fn as_str(self) -> &'static str {
        match self {
            Self::Ascending => "ascending",
            Self::Descending => "descending",
        }
    }
}

impl<'a> Column<'a> {
    /// A column that absorbs slack and drops after the optional ones.
    #[must_use]
    pub const fn new(name: &'a str) -> Self {
        Self {
            name,
            width: Width::Fill,
            priority: Priority::Secondary,
            sortable: false,
            sorted: None,
        }
    }

    /// Whether this column survives at the given cutoff.
    ///
    /// A renderer narrows by raising the cutoff, and never by counting
    /// positions.
    #[must_use]
    pub const fn kept_at(&self, cutoff: Priority) -> bool {
        (self.priority as u8) >= (cutoff as u8)
    }
}

/// What a table cell holds.
///
/// [`RowPart`] for tables, and it exists for the same reason: a part that
/// carries a control is not text, and a renderer with one class for the whole
/// cell paints it as though it were. `makeover-webview` emitted a single
/// `.cell` until 0.25.0, so a button in a cell inherited the cell's content
/// colour, which is the exact drift [`RowPart::intent`] prevents for rows and
/// prevented for nothing here.
///
/// Four members, and the count is what quasi's `Cell` was measured to carry:
/// a value, tokens (33 cells across 22 server templates), actions (30 rows
/// carrying a control, 5 beside a value) and a link (35 cells across 18
/// templates). Nothing was added past what something holds.
///
/// `#[non_exhaustive]` for [`RowPart`]'s reason: growth here must not be a
/// lockstep event across three renderers.
///
/// # No hover-reveal
///
/// [`RowPart`] carried a `revealed_on_hover` until 0.13.0 retired it, and this
/// enum never gets one. A cell's actions are shown at rest in every consumer
/// measured, and a member nothing uses is one three renderers owe an answer
/// for.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub enum CellPart {
    /// The cell's own text.
    Value,
    /// Small labelled things in the cell: a status badge, a chip.
    Tokens,
    /// Controls that act on what the row is about.
    Actions,
    /// The cell's value, where the value is itself a link.
    Link,
}

impl CellPart {
    /// The content intent the part takes.
    ///
    /// One part is text and three are not, so three answer with the intent
    /// inheriting already gives. That is [`RowPart::intent`]'s shape with the
    /// text side narrower: a cell's secondary and muted readings are the
    /// column's business, not the cell's.
    #[must_use]
    pub const fn intent(self) -> &'static str {
        match self {
            Self::Value => "content",
            // A token carries its own tone, and a part-level intent underneath
            // it would fight the token sitting on it.
            Self::Tokens => "content",
            // Actions carry controls rather than text.
            Self::Actions => "content",
            // A link takes the action colour from the control it is, rather
            // than the cell's text colour from the cell it sits in.
            Self::Link => "content",
        }
    }
}

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

    #[test]
    fn the_four_readiness_states_are_one_axis_and_only_one_shows_content() {
        // Mutually exclusive is the test for one enum against several fields: a
        // region shows its content, or that it is coming, or that there is none,
        // or that it broke. Never two.
        assert!(Readiness::Ready.shows_content());
        for state in [Readiness::Pending, Readiness::Empty, Readiness::Failed] {
            assert!(!state.shows_content());
        }
    }

    #[test]
    fn a_region_shows_all_of_its_children_unless_it_says_otherwise() {
        // The default is the behaviour every region had before this member
        // existed, which is what keeps it additive: a description written
        // against 0.22.0 says the same thing under 0.23.0.
        assert_eq!(Showing::default(), Showing::All);
        assert!(!Showing::All.selective());
    }

    #[test]
    fn only_a_disclosure_can_show_nothing() {
        // The two derived idioms differ in one respect and this is it. A
        // carousel's row moves between frames and never reaches empty; a
        // disclosure's summary line is the same control wearing its closed
        // state, so a renderer has to know which it is drawing.
        assert!(Showing::AtMostOne.dismissible());
        assert!(!Showing::One.dismissible());
        assert!(!Showing::All.dismissible());

        // Both are selective, though. Deriving chrome is one question and
        // whether that chrome closes is another.
        assert!(Showing::One.selective());
        assert!(Showing::AtMostOne.selective());
    }

    #[test]
    fn an_empty_region_is_not_a_broken_one() {
        // An empty list is the normal state of a new install. Drawing it in a
        // danger tone reports a fault where there is none, and this is the one
        // place the distinction is carried.
        assert_eq!(Readiness::Empty.tone(), Tone::Neutral);
        assert_eq!(Readiness::Failed.tone(), Tone::Danger);
        assert_eq!(Readiness::Pending.tone(), Tone::Neutral);
    }

    #[test]
    fn a_column_can_be_sorted_without_being_sortable() {
        // Both combinations mean something, which is why the two fields are
        // independent rather than one implying the other. A list ordered by a
        // key the user cannot change is a real thing with a caret worth drawing.
        let fixed = Column {
            sorted: Some(Sort::Descending),
            ..Column::new("Created")
        };

        assert!(!fixed.sortable);
        assert_eq!(fixed.sorted.map(Sort::as_str), Some("descending"));

        let offered = Column {
            sortable: true,
            ..Column::new("Name")
        };
        assert_eq!(offered.sorted, None);
    }

    #[test]
    fn a_direction_flips_and_says_what_it_is() {
        assert_eq!(Sort::Ascending.reversed(), Sort::Descending);
        assert_eq!(Sort::Descending.reversed().reversed(), Sort::Descending);
        assert_eq!(Sort::Ascending.as_str(), "ascending");
    }

    #[test]
    fn a_figure_carries_its_tone_because_no_renderer_can_derive_it() {
        // Three of goingson's five sites tone the figure by their own means, so
        // tone is carried at every site that needs it and derived at none. The
        // same reasoning `Meter` reached, from a different direction.
        let streak = Figure::new("0", "Current Streak").tone(Tone::Warning);
        assert_eq!(streak.tone, Tone::Warning);
        assert_eq!(Figure::new("17", "Total").tone, Tone::Neutral);
    }

    #[test]
    fn a_figures_change_is_the_toned_part_and_is_absent_by_default() {
        // 0.13.0. The MNW server's stat card is a label, a value and a delta,
        // across four screens, and the delta is what reads as good or bad. Tone
        // had no consumer before this: the figure itself is an ordinary fact.
        let views = Figure::new("1,204", "Views")
            .change("+12.5%")
            .tone(Tone::Success);
        assert_eq!(views.change, Some("+12.5%"));
        assert_eq!(views.tone, Tone::Success);

        // A figure with nothing to compare against says so by having no change,
        // rather than by carrying an empty string a renderer has to test for.
        assert_eq!(Figure::new("3.1%", "Conversion").change, None);
    }

    #[test]
    fn a_figures_value_is_text_because_only_the_app_knows_what_it_is() {
        // "84%", "12/30", "3d". A figure is whatever the app computed, already
        // formatted, and that is the line between this and `Meter`: a meter is
        // a proportion a renderer draws, a figure is a fact it sets in type.
        for value in ["84%", "12/30", "3d"] {
            assert_eq!(Figure::new(value, "Rate").value, value);
        }
    }

    #[test]
    fn a_proportion_is_a_row_part_and_takes_no_intent_of_its_own() {
        // The meter carries the tone, so a part-level intent underneath would
        // fight it. Same answer `Tokens` needed, for the same reason.
        assert_eq!(RowPart::Proportion.intent(), RowPart::Tokens.intent());
    }

    #[test]
    fn a_file_field_is_drawn_and_offers_no_options() {
        // It is a control the user operates, unlike `Hidden`, and it does not
        // pick from a list the description carries, unlike `Select`.
        assert!(FieldKind::File.visible());
        assert!(!FieldKind::File.offers_options());
        assert!(!FieldKind::File.confidential());
    }

    #[test]
    fn a_constraint_is_a_fact_about_the_question_and_not_a_verdict() {
        // The whole model: the description carries the rule, the renderer emits
        // its host's idiom, and `error` is what arrives back when someone
        // validated. Nothing here decides a value is wrong.
        let field = Field {
            max_length: Some(100),
            min: Some("1"),
            max: Some("240"),
            required: true,
            ..Field::new(FieldKind::Number, "minutes", "Minutes")
        };
        assert!(!field.invalid());

        // A bound is text because it is only a number for some of the kinds
        // that take one. goingson has both shapes live.
        let when = Field {
            min: Some("2026-08-09T14:30"),
            ..Field::new(FieldKind::Text, "starts", "Starts")
        };
        assert_eq!(when.min, Some("2026-08-09T14:30"));
    }

    #[test]
    fn a_meter_keeps_the_over_run_the_percentage_throws_away() {
        // The whole reason this is a pair. goingson's `Task::time_progress`
        // clamps to 100 and then carries `is_over_estimate` beside it to say
        // what the clamp dropped; a meter says both from one fact.
        let over = Meter::new(45, 30);
        assert_eq!(over.percent(), 100);
        assert!(over.overflowing());

        let exact = Meter::new(30, 30);
        assert_eq!(exact.percent(), over.percent());
        assert!(!exact.overflowing());
    }

    #[test]
    fn an_empty_set_does_not_divide_by_zero() {
        // Sayable on purpose, so it has to be answerable. A meter over an
        // unloaded count is what an app actually has for a frame.
        let none = Meter::new(0, 0);
        assert_eq!(none.percent(), 0);
        assert!(none.is_empty());
        assert!(!none.overflowing());
    }

    #[test]
    fn the_ratio_survives_where_a_percentage_would_not() {
        // Given 43 nothing can recover "3 of 7", which is why the numbers are
        // carried and the label names only the noun.
        let m = Meter::new(3, 7).label("subtasks");
        assert_eq!(m.percent(), 42);
        assert_eq!((m.done, m.total), (3, 7));
        assert_eq!(m.label, Some("subtasks"));
    }

    #[test]
    fn tone_is_carried_because_no_renderer_can_derive_it() {
        // The same fullness means opposite things on two of goingson's bars,
        // and only the app knows which.
        let subtasks = Meter::new(9, 10).tone(Tone::Success);
        let estimate = Meter::new(9, 10).tone(Tone::Danger);
        assert_eq!(subtasks.percent(), estimate.percent());
        assert_ne!(subtasks.tone, estimate.tone);
        // Untoned by default: a bar says nothing about status until something
        // says so, the same way a row is not selectable until told.
        assert_eq!(Meter::new(9, 10).tone, Tone::Neutral);
    }

    #[test]
    fn an_act_is_reachable_until_it_is_disabled() {
        // The one member a renderer must branch on, and since 0.19.0 the only
        // member there is. A stated state is not by itself a reason to stop
        // answering, which is the distinction `State` makes and every
        // hand-rolled button in the tree had to remember.
        assert!(!Act::new("Save").disabled());
        assert!(Act::new("Save").state(State::Disabled).disabled());
    }

    #[test]
    fn an_act_carries_its_key_because_a_terminal_has_nothing_else() {
        // No key is the ordinary case, and the webview hosts that ignore it
        // are why it stayed optional.
        assert_eq!(Act::new("Delete").key, None);
        let quit = Act::new("Quit").key("q").tone(Tone::Danger);
        assert_eq!(quit.key, Some("q"));
        assert_eq!(quit.tone, Tone::Danger);
    }

    #[test]
    fn a_meter_does_not_overflow_on_large_counts() {
        // done * 100 in u32 would wrap somewhere past 42 million. Counts that
        // size are not tasks, but a description layer that silently reports 3%
        // for a full bar is worse than one that is slow.
        let big = Meter::new(u32::MAX, u32::MAX);
        assert_eq!(big.percent(), 100);
        assert!(!big.overflowing());
    }

    #[test]
    fn inset_is_raised_with_the_light_moved() {
        let (rl, rd) = Bevel::Raised.edges();
        let (il, id) = Bevel::Inset.edges();
        assert_eq!((rl, rd), (Edge::Light, Edge::Dark));
        assert_eq!((il, id), (rd, rl));
    }

    #[test]
    fn pressing_twice_is_a_no_op() {
        for b in [Bevel::Raised, Bevel::Inset] {
            assert_eq!(b.pressed().pressed(), b);
        }
    }

    #[test]
    fn a_raised_region_is_never_filled_with_a_recessed_surface() {
        // The bug this vocabulary exists to make unrepresentable.
        assert_eq!(Depth::Raised.fill(), Some(Fill::Raised));
        assert_eq!(Depth::Raised.bevel(), Some(Bevel::Raised));
        assert_eq!(Depth::Well.bevel(), Some(Bevel::Inset));
        assert_ne!(Depth::Well.fill(), Depth::Raised.fill());
    }

    #[test]
    fn state_is_orthogonal_to_depth() {
        // The reason State is its own axis and not a Depth member: a disabled
        // button and a disabled field are both disabled and are not the same
        // shape, which one shared variant could not have said.
        assert_eq!(Depth::Raised.fill(), Some(Fill::Raised));
        assert_eq!(Depth::Well.fill(), Some(Fill::Well));
        assert!(State::Disabled.suppresses_interaction());
    }

    #[test]
    fn only_disabled_stops_answering() {
        // Kept in spirit from the version where `Focus` was the counter-example:
        // suppressing interaction is `Disabled`'s alone, so a member added here
        // later does not get to inherit it by being a state.
        assert!(State::Disabled.suppresses_interaction());
    }

    #[test]
    fn disabled_resolves_against_an_intent_makeover_already_derives() {
        // No new token, so this costs no `makeover` release.
        assert_eq!(State::Disabled.token(), "content-muted");
    }

    #[test]
    fn flat_has_neither_edge_nor_fill() {
        assert_eq!(Depth::Flat.bevel(), None);
        assert_eq!(Depth::Flat.fill(), None);
    }

    #[test]
    fn sunken_is_recessed_by_colour_with_no_edge() {
        // The one member carrying a fill without a bevel. A renderer that
        // assumes the two arrive together drops the fill silently, which is
        // exactly what makeover-webview did before 0.3.0.
        assert_eq!(Depth::Sunken.fill(), Some(Fill::Sunken));
        assert_eq!(Depth::Sunken.bevel(), None);
    }

    #[test]
    fn sunken_and_flat_are_different_claims() {
        // Both edgeless, and only one of them needs a colour. Collapsing them
        // is what left an unchosen tab unsayable.
        assert_eq!(Depth::Flat.bevel(), Depth::Sunken.bevel());
        assert_ne!(Depth::Flat.fill(), Depth::Sunken.fill());
    }

    #[test]
    fn a_sunken_surface_is_not_a_well() {
        // Authored in opposite directions: makeover derives surface-well by
        // inverting against the theme's content colour, while surface-sunken is
        // authored and may sit darker than raised.
        assert_ne!(Fill::Sunken, Fill::Well);
        assert_eq!(Fill::Sunken.token(), "surface-sunken");
        assert_eq!(Fill::Well.token(), "surface-well");
    }

    #[test]
    fn every_selector_describes_both_of_its_states() {
        // The gap 0.3.0 closed. Before it, only `chosen` existed and the
        // unchosen option fell through to Flat at every renderer.
        for s in [Selector::Tabs, Selector::Segmented, Selector::Toggle] {
            assert_ne!(
                s.chosen(),
                s.unchosen(),
                "{s:?} cannot tell picked from unpicked"
            );
        }
    }

    #[test]
    fn only_a_tab_inverts_the_other_way() {
        // Tabs recede so the chosen one comes forward; a segment and a toggle
        // stand up so the chosen one is held in. That inversion is the whole
        // content of "picked" once colour is deferred, and it is why the three
        // are not one member with a flag.
        assert_eq!(Selector::Tabs.unchosen(), Depth::Sunken);
        assert_eq!(Selector::Tabs.chosen(), Depth::Raised);

        for s in [Selector::Segmented, Selector::Toggle] {
            assert_eq!(s.unchosen(), Depth::Raised);
            assert_eq!(s.chosen(), Depth::Well);
            // Held in is what pressing produces: one appearance, two reasons.
            assert_eq!(s.unchosen().pressed(), s.chosen());
        }
    }

    #[test]
    fn pressing_a_card_makes_a_well() {
        assert_eq!(Depth::Raised.pressed(), Depth::Well);
        assert_eq!(
            Depth::Raised.pressed().bevel(),
            Depth::Raised.bevel().map(Bevel::pressed)
        );
        // Only raised regions respond to being pressed.
        assert_eq!(Depth::Flat.pressed(), Depth::Flat);
        assert_eq!(Depth::Well.pressed(), Depth::Well);
        // An overlay is a surface, not a control.
        assert_eq!(Depth::Overlay.pressed(), Depth::Overlay);
    }

    #[test]
    fn an_overlay_is_lifted_rather_than_edged() {
        // The wave-2 rule: a surface over the page takes elevation, a surface
        // in the page takes a bevel. Both halves come off the one Depth, so
        // they cannot disagree.
        assert_eq!(Depth::Overlay.fill(), Some(Fill::Overlay));
        assert_eq!(Depth::Overlay.bevel(), None);

        // Three depths have no bevel and they are not the same claim. Flat has
        // nothing to separate from, Sunken's colour is doing the separating,
        // and an overlay is separated by the lift.
        assert_ne!(Depth::Overlay.fill(), Depth::Sunken.fill());
        assert_ne!(Depth::Overlay.fill(), Depth::Flat.fill());
    }

    #[test]
    fn intents_name_makeover_tokens_and_nothing_else() {
        assert_eq!(Edge::Light.token(), "bevel-light");
        assert_eq!(Edge::Dark.token(), "bevel-dark");
        assert_eq!(Fill::Raised.token(), "surface-raised");
        assert_eq!(Fill::Well.token(), "surface-well");
        // No value ever leaves this crate.
        for t in [
            Edge::Light.token(),
            Edge::Dark.token(),
            Tone::Danger.token(),
            Tone::Neutral.token(),
            State::Disabled.token(),
        ] {
            assert!(!t.starts_with('#'), "{t} looks like a value");
            assert!(
                !t.chars().next().unwrap().is_ascii_digit(),
                "{t} is a value"
            );
        }
    }

    #[test]
    fn a_badge_cannot_be_pressed_and_a_chip_latches() {
        // The one line that runs through all three apps' taxonomies.
        assert!(!Token::Badge.interactive());
        assert!(Token::Chip { removable: false }.interactive());
        assert!(Token::Chip { removable: true }.interactive());

        // A badge is a label, so giving it an edge would lie about it.
        assert_eq!(Token::Badge.depth(false), Depth::Flat);
        assert_eq!(Token::Badge.depth(true), Depth::Flat);

        // A latched chip wears the same shape a pressed one does.
        let chip = Token::Chip { removable: false };
        assert_eq!(chip.depth(false), Depth::Raised);
        assert_eq!(chip.depth(true), Depth::Raised.pressed());
    }

    #[test]
    fn a_toast_and_a_banner_differ_in_more_than_placement() {
        assert!(Notice::Toast.transient());
        assert!(!Notice::Banner.transient());
        // A toast floats above the page; a banner rests in the flow.
        assert_eq!(Notice::Toast.fill(), Fill::Overlay);
        assert_eq!(Notice::Banner.fill(), Fill::Raised);
    }

    #[test]
    fn emphasis_falls_off_down_the_row() {
        // `revealed_on_hover` was asserted here until 0.13.0 retired it. It said
        // a row's actions stay hidden until hover, which stopped being true when
        // makeover-webview 0.23.0 showed them at rest, and nothing had consumed
        // it for a release either way.
        assert_eq!(RowPart::Primary.intent(), "content");
        assert_eq!(RowPart::Secondary.intent(), "content-secondary");
        assert_eq!(RowPart::Meta.intent(), "content-muted");
    }

    #[test]
    fn a_token_part_carries_no_intent_of_its_own() {
        // Each token carries its own tone, so a part-level intent underneath
        // would fight the thing sitting on it. Same reasoning as actions, which
        // is why they answer alike.
        assert_eq!(RowPart::Tokens.intent(), RowPart::Actions.intent());
        assert_eq!(RowPart::Tokens.intent(), "content");
    }

    #[test]
    fn the_two_temporal_kinds_are_the_two_that_name_a_moment() {
        // The pair is named once so a host with parsing to do asks here rather
        // than spelling it out, which is `offers_options`' reason.
        assert!(FieldKind::Date.temporal());
        assert!(FieldKind::DateTime.temporal());

        for kind in [
            FieldKind::Text,
            FieldKind::Secret,
            FieldKind::Number,
            FieldKind::Email,
            FieldKind::Url,
            FieldKind::Tel,
            FieldKind::Textarea,
            FieldKind::Select,
            FieldKind::Radio,
            FieldKind::Checkbox,
            FieldKind::File,
            FieldKind::Hidden,
        ] {
            assert!(!kind.temporal(), "{kind:?}");
        }
    }

    #[test]
    fn a_date_carries_no_time_and_a_datetime_carries_no_zone() {
        // The formats are the whole reason the members are worth naming apart
        // from text, so the doc comments and the constants have to agree. A
        // host reading one and meeting the other is the silent failure.
        assert_eq!(DATE_FORMAT, "%Y-%m-%d");
        assert!(!DATE_FORMAT.contains("%H"), "a day carries no hour");

        assert_eq!(DATETIME_FORMAT, "%Y-%m-%dT%H:%M");
        assert!(
            DATETIME_FORMAT.starts_with(DATE_FORMAT),
            "a moment starts with the day it is on"
        );
        // Local, and that is a property of the value rather than an omission.
        assert!(!DATETIME_FORMAT.contains("%Z"), "no zone name");
        assert!(!DATETIME_FORMAT.ends_with('Z'), "not UTC-stamped");
        assert!(!DATETIME_FORMAT.contains("%S"), "no seconds by default");
    }

    #[test]
    fn a_temporal_kind_takes_a_label_above_it_and_offers_no_options() {
        // Neither is a checkbox and neither is a fixed set, so both fall where
        // text does. Asserted because a new kind lands in three predicates and
        // only one of them is the interesting one.
        for kind in [FieldKind::Date, FieldKind::DateTime] {
            assert!(kind.visible(), "{kind:?}");
            assert!(!kind.confidential(), "{kind:?}");
            assert!(!kind.labels_itself(), "{kind:?}");
            assert!(!kind.offers_options(), "{kind:?}");
        }
    }

    #[test]
    fn a_cell_part_names_an_intent_and_only_the_value_is_text() {
        // The table half of what RowPart::intent does for rows. A cell holding
        // a control and a cell holding text answered alike until 0.14.0, and a
        // control in a cell took the cell's text colour.
        assert_eq!(CellPart::Value.intent(), "content");

        for part in [CellPart::Tokens, CellPart::Actions, CellPart::Link] {
            // Each for its own reason -- a token carries its tone, an action is
            // a control, a link takes the action colour -- and all three reach
            // the intent inheriting already gives.
            assert_eq!(part.intent(), CellPart::Value.intent(), "{part:?}");
        }
    }

    #[test]
    fn every_cell_part_answers_with_a_token_and_never_a_value() {
        for part in [
            CellPart::Value,
            CellPart::Tokens,
            CellPart::Actions,
            CellPart::Link,
        ] {
            let intent = part.intent();
            assert!(!intent.is_empty(), "{part:?} names nothing");
            assert!(!intent.starts_with('#'), "{part:?} looks like a value");
        }
    }

    #[test]
    fn a_separator_is_what_tells_a_section_from_a_subsection() {
        assert!(Heading::Section.separated());
        assert!(!Heading::Subsection.separated());
        assert!(!Heading::Page.separated());
    }

    #[test]
    fn a_chosen_segment_is_held_in_and_a_chosen_tab_comes_forward() {
        assert_eq!(Selector::Segmented.chosen(), Depth::Well);
        assert_eq!(Selector::Toggle.chosen(), Depth::Well);
        // The exception, and the whole folder semantic: the open tab joins its
        // pane rather than sinking away from it.
        assert_eq!(Selector::Tabs.chosen(), Depth::Raised);

        // A held-in segment is indistinguishable from a pressed raised one,
        // which is the economy the light model buys over a colour swap.
        assert_eq!(Selector::Segmented.chosen(), Depth::Raised.pressed());

        // A toggle stands alone; the other two are built out of parts that
        // touch.
        assert!(Selector::Segmented.abutting());
        assert!(Selector::Tabs.abutting());
        assert!(!Selector::Toggle.abutting());
    }

    #[test]
    fn a_pane_is_looked_into_and_a_band_is_not() {
        assert_eq!(Region::Pane.depth(), Depth::Well);
        assert_eq!(Region::Modal.depth(), Depth::Raised);
        for r in [
            Region::Band,
            Region::Sidebar,
            Region::Split,
            Region::TabGroup,
        ] {
            assert_eq!(r.depth(), Depth::Flat, "{r:?} should carry no edge");
        }
    }

    #[test]
    fn exactly_one_region_is_opaque() {
        // The escape hatch is one member and stays one member. If a second
        // undescribed region ever appears, the description has started
        // conceding rather than deferring.
        for r in [
            Region::Band,
            Region::Sidebar,
            Region::Pane,
            Region::Split,
            Region::TabGroup,
            Region::Modal,
            // A widget is described, and that is the whole of what separates it
            // from a bespoke here. Both carry a name this crate never reads;
            // only one of them has contents under it that a renderer which does
            // not know the name can still walk.
            Region::Widget { name: "carousel" },
        ] {
            assert!(r.described(), "{r:?} should be describable");
        }
        assert!(!Region::Bespoke { name: "day-plan" }.described());
    }

    #[test]
    fn a_picture_that_says_nothing_is_a_claim_and_not_an_oversight() {
        // The distinction a renderer with no graphics protocol runs on: draw
        // the words, or draw nothing. Standing in for a decorative rule with
        // the word "decoration" is worse than leaving the space empty.
        assert!(Image::new("The library view, mid-import").speaks());
        assert!(!Image::new("").speaks());
    }

    #[test]
    fn a_caption_and_alt_text_are_not_the_same_line() {
        // A caption is content everybody reads; alt text stands in for the
        // picture. A screenshot with a caption still needs alt text.
        let shot = Image::new("A file list with three rows selected").caption("The library view");
        assert_eq!(shot.caption, Some("The library view"));
        assert!(shot.speaks());
        assert_ne!(shot.alt, shot.caption.unwrap());
    }

    #[test]
    fn a_picture_can_say_how_much_room_to_hold() {
        // The whole point: a renderer reserves from the ratio, so the space is
        // right at any width. A fixed height would only be right at one.
        let shot = Image::new("a screenshot").intrinsic(5120, 3412);
        let e = shot.intrinsic.expect("carried");
        assert_eq!((e.width, e.height), (5120, 3412));
        assert!((e.ratio().unwrap() - 1.5006).abs() < 0.001);
    }

    #[test]
    fn a_picture_with_no_dimensions_reserves_nothing_rather_than_guessing() {
        // `None` is honest: a creator upload whose size was never recorded does
        // not know it. A renderer must not invent one.
        assert_eq!(Image::new("unknown upload").intrinsic, None);
        assert_eq!(Extent::new(0, 10).ratio(), None);
        assert_eq!(Extent::new(10, 0).ratio(), None);
    }

    #[test]
    fn a_picture_is_wanted_now_unless_the_app_says_otherwise() {
        // Eager is the safe default and lazy is the opt-in, because deferring
        // something already on screen saves nothing and moves its shift later.
        assert_eq!(Image::new("hero").loading, Loading::Eager);
        assert_eq!(Loading::default(), Loading::Eager);
        assert_eq!(Image::new("frame 2").lazy().loading, Loading::Lazy);
    }

    #[test]
    fn a_picture_keeps_its_own_proportions_unless_told_otherwise() {
        // The default is the one that shows the whole picture at its own shape,
        // so a renderer ignoring Fit entirely is still right about the common
        // case. The shipped MNW carousel sets no object-fit at all, which is
        // this.
        assert_eq!(Image::new("a").fit, Fit::Natural);
        assert_eq!(Fit::default(), Fit::Natural);
        assert_eq!(Image::new("a").fit(Fit::Cover).fit, Fit::Cover);
    }

    #[test]
    fn a_widget_inherits_its_depth_the_way_a_bespoke_does() {
        // Stronger than the bespoke case: a widget is drawn by whichever
        // renderer recognises the name, so a depth chosen here would be this
        // crate deciding a carousel is raised on every host.
        assert_eq!(Region::Widget { name: "carousel" }.depth(), Depth::Flat);
        assert_eq!(Region::Widget { name: "pager" }.depth(), Depth::Flat);
    }

    #[test]
    fn a_name_is_readable_without_asking_which_member_carried_it() {
        // A renderer dispatching on a name wants the string, not the member.
        // Writing that `matches!` at each renderer is how the two drift apart.
        assert_eq!(Region::Widget { name: "carousel" }.name(), Some("carousel"));
        assert_eq!(
            Region::Bespoke { name: "day-plan" }.name(),
            Some("day-plan")
        );

        for r in [
            Region::Band,
            Region::Sidebar,
            Region::Pane,
            Region::Split,
            Region::TabGroup,
            Region::Modal,
        ] {
            assert_eq!(r.name(), None, "{r:?} names nothing an app chose");
        }
    }

    #[test]
    fn a_bespoke_region_inherits_its_depth_rather_than_choosing_one() {
        // The app owns the contents, not the placement. An app that wants its
        // timeline in a well frames it in a Pane.
        assert_eq!(Region::Bespoke { name: "day-plan" }.depth(), Depth::Flat);
        assert_eq!(Region::Bespoke { name: "kanban" }.depth(), Depth::Flat);
    }

    #[test]
    fn a_screen_with_a_bespoke_region_is_still_a_whole_screen() {
        // The argument the member exists for: goingson's day-plan has to be
        // routable, or the description covers only the boring screens and the
        // interesting four need a second path beside the router.
        let day_plan = [
            Region::Band,
            Region::Bespoke { name: "day-plan" },
            Region::Sidebar,
        ];
        assert_eq!(day_plan.iter().filter(|r| r.described()).count(), 2);
        assert_eq!(day_plan.iter().filter(|r| !r.described()).count(), 1);
    }

    #[test]
    fn a_secret_field_is_marked_as_one_and_a_hidden_field_is_not_drawn() {
        let secret = Field::new(FieldKind::Secret, "password", "Password");
        assert!(secret.kind.confidential());
        assert!(secret.kind.visible());

        assert!(!FieldKind::Hidden.visible());
        // Nothing else is confidential, or the marker means nothing.
        for k in [
            FieldKind::Text,
            FieldKind::Number,
            FieldKind::Textarea,
            FieldKind::Select,
            FieldKind::Checkbox,
            FieldKind::Hidden,
        ] {
            assert!(!k.confidential(), "{k:?} should not be confidential");
        }

        // Only a checkbox carries its own label.
        assert!(FieldKind::Checkbox.labels_itself());
        assert!(!FieldKind::Text.labels_itself());
    }

    #[test]
    fn a_plain_field_offers_nothing_and_a_select_offers_its_options() {
        let text = Field::new(FieldKind::Text, "title", "Title");
        assert!(text.options.is_empty());
        assert_eq!(text.placeholder, None);

        let sizes = [Choice::plain("small"), Choice::plain("large")];
        let select = Field::select("size", "Size", &sizes);
        assert_eq!(select.kind, FieldKind::Select);
        assert_eq!(select.options.len(), 2);
    }

    #[test]
    fn a_choice_says_what_submits_and_what_is_read_apart() {
        // The whole reason it is two strings. `plain` is the case where they
        // coincide, and it is a shorthand rather than the general shape.
        let plain = Choice::plain("7");
        assert_eq!((plain.value, plain.label), ("7", "7"));

        let spelled = Choice {
            value: "7",
            label: "One week",
        };
        assert_ne!(spelled.value, spelled.label);
    }

    #[test]
    fn a_radio_asks_the_same_question_as_a_select_and_is_not_the_same_kind() {
        // Both offer a fixed set and both read `options`, so the two
        // constructors differ in exactly one thing. That one thing is the
        // point: a renderer decides whether the alternatives are readable
        // without opening anything, and it can only decide that if the
        // description said which question was asked.
        let styles = [
            Choice {
                value: "copy",
                label: "Copy samples in",
            },
            Choice {
                value: "reference",
                label: "Reference in place",
            },
        ];
        let radio = Field::radio("storage", "Storage style", &styles);
        let select = Field::select("storage", "Storage style", &styles);

        assert_eq!(radio.kind, FieldKind::Radio);
        assert_ne!(radio.kind, select.kind);
        assert_eq!(radio.options, select.options);
        assert_eq!(
            Field {
                kind: select.kind,
                ..radio
            },
            select
        );
    }

    #[test]
    fn exactly_the_option_taking_kinds_say_so() {
        // The renderers branch on this rather than on a list of their own, so
        // a kind added without a decision here renders its options nowhere.
        assert!(FieldKind::Select.offers_options());
        assert!(FieldKind::Radio.offers_options());
        for kind in [
            FieldKind::Text,
            FieldKind::Secret,
            FieldKind::Number,
            FieldKind::Email,
            FieldKind::Url,
            FieldKind::Tel,
            FieldKind::Textarea,
            FieldKind::Checkbox,
            FieldKind::Hidden,
        ] {
            assert!(!kind.offers_options(), "{kind:?} does not offer options");
        }
    }

    #[test]
    fn a_radio_group_takes_a_label_even_though_its_options_carry_their_own() {
        // The near-miss: each option is labelled beside its own button, so a
        // renderer could plausibly read the group as self-labelling and drop
        // the question. Checkbox is the only kind that does that.
        assert!(!FieldKind::Radio.labels_itself());
        assert!(FieldKind::Checkbox.labels_itself());
    }

    #[test]
    fn a_select_with_no_options_is_sayable() {
        // An app whose option list has not loaded has exactly this. Making it
        // unrepresentable would push the state somewhere less visible, and a
        // renderer drawing an empty select reports it on screen.
        let loading = Field::select("project", "Project", &[]);
        assert!(loading.options.is_empty());
    }

    #[test]
    fn the_description_carries_the_question_and_never_the_answer() {
        // The line 0.8.0 drew. Placeholder and options are properties of what
        // is being asked; the current value is what came back, and no field
        // here holds one.
        let f = Field {
            placeholder: Some("yyyy-mm-dd"),
            ..Field::new(FieldKind::Text, "due", "Due")
        };
        assert_eq!(f.placeholder, Some("yyyy-mm-dd"));
        // A placeholder is not a label, and having one does not excuse the
        // field from carrying the other.
        assert_eq!(f.label, "Due");
    }

    #[test]
    fn a_field_reports_its_own_error_state() {
        let mut f = Field::new(FieldKind::Text, "title", "Title");
        assert!(!f.invalid());
        f.error = Some("Required");
        assert!(f.invalid());
    }

    #[test]
    fn columns_drop_by_priority_and_never_by_position() {
        let cols = [
            Column {
                width: Width::Fill,
                priority: Priority::Essential,
                ..Column::new("Title")
            },
            Column {
                width: Width::Fixed,
                priority: Priority::Secondary,
                ..Column::new("Due")
            },
            Column {
                width: Width::Fixed,
                priority: Priority::Optional,
                ..Column::new("Estimate")
            },
        ];

        // Widest: everything survives.
        assert_eq!(
            cols.iter()
                .filter(|c| c.kept_at(Priority::Optional))
                .count(),
            3
        );
        // Narrower: the optional column goes first.
        let kept: Vec<_> = cols
            .iter()
            .filter(|c| c.kept_at(Priority::Secondary))
            .map(|c| c.name)
            .collect();
        assert_eq!(kept, ["Title", "Due"]);
        // Narrowest: only what identifies the row.
        let kept: Vec<_> = cols
            .iter()
            .filter(|c| c.kept_at(Priority::Essential))
            .map(|c| c.name)
            .collect();
        assert_eq!(kept, ["Title"]);
    }

    #[test]
    fn inserting_a_column_does_not_move_what_gets_dropped() {
        // The bug the ordinal form has and this form cannot: goingson hides
        // `nth-child(n+5)` against a seven-column table, so a column inserted
        // anywhere to the left silently hides a different one.
        let before = [
            Column::new("Title"),
            Column {
                width: Width::Fixed,
                priority: Priority::Optional,
                ..Column::new("Estimate")
            },
        ];
        let after = [
            Column::new("Title"),
            Column::new("Project"), // inserted
            Column {
                width: Width::Fixed,
                priority: Priority::Optional,
                ..Column::new("Estimate")
            },
        ];

        fn dropped<'a>(cols: &[Column<'a>]) -> Vec<&'a str> {
            cols.iter()
                .filter(|c| !c.kept_at(Priority::Secondary))
                .map(|c| c.name)
                .collect()
        }
        assert_eq!(dropped(&before), ["Estimate"]);
        assert_eq!(dropped(&after), ["Estimate"]);
    }

    #[test]
    fn an_arrangement_carries_the_tab_group_as_a_modifier() {
        // goingson uses the tab group inside the content region rather than
        // instead of one, so it is not a third arrangement.
        let go = Arrangement::list_detail(true);
        let plain = Arrangement::list_detail(false);
        assert_ne!(go, plain);
        assert_ne!(go, Arrangement::sidebar_content());
    }

    #[test]
    fn a_share_is_a_proportion_and_resolves_the_same_way_everywhere() {
        // The point of the member: a terminal reading columns and a webview
        // reading a grid honour one fact, so two hosts showing one screen agree
        // about its proportions.
        assert_eq!(Share::LIST.as_percent(), 40);
        assert_eq!(Share::LIST.of(100), 40);
        assert_eq!(
            Share::SIDEBAR.of(96),
            24,
            "quasi-tui's 24 columns, said as a quarter"
        );
    }

    #[test]
    fn a_region_never_resolves_to_nothing() {
        // A region the description named should be visible. A zero-width one
        // reads on screen as a region that vanished, which is the hardest kind
        // of bug to find from what is drawn.
        assert_eq!(Share::percent(5).of(1), 1);
        assert_eq!(Share::percent(5).of(0), 1);
    }

    #[test]
    fn a_share_outside_the_range_is_clamped_rather_than_refused() {
        assert_eq!(Share::percent(0), Share::percent(5));
        assert_eq!(Share::percent(200), Share::percent(95));
    }

    #[test]
    fn the_share_rides_on_the_arrangement_that_knows_which_question_it_is() {
        // How much a sidebar takes and how much a list side takes are different
        // questions, and this enum is the only thing that knows which is being
        // asked.
        assert_eq!(Arrangement::sidebar_content().share(), Share::SIDEBAR);
        assert_eq!(Arrangement::list_detail(false).share(), Share::LIST);

        let narrow = Arrangement::sidebar_content().with_share(Share::percent(20));
        assert_eq!(narrow.share(), Share::percent(20));
        assert!(matches!(narrow, Arrangement::SidebarContent { .. }));
    }

    #[test]
    fn a_measure_defaults_to_the_one_53_of_69_templates_asked_for() {
        // The default is meaningful: a screen nobody said anything about uses
        // the window it was given.
        assert_eq!(Measure::default(), Measure::Wide);
        assert_eq!(Measure::Reading.as_str(), "reading");
    }

    #[test]
    fn readiness_names_the_state_and_not_the_shimmer() {
        // Two members and no third. If a skeleton ever appears in this enum,
        // the deferral rule has been broken.
        assert_ne!(Readiness::Ready, Readiness::Pending);
    }
}