makeover_webview/

form.rs

//! Phase B, the forms half: the escaper, the field classes and the rules a
//! described field's markup is drawn with.
//!
//! The field markup itself is drawn by the quasi hosts (`quasi-webview` for
//! this renderer), which own the vocabulary member. What stays here is what
//! every caller writing HTML needs and what the generated stylesheet has to
//! agree with: [`escape`] and [`escape_into`], [`Markup`] and [`Filling`], the
//! class lists, and the rules behind them.
//!
//! # Why one escaper is enough
//!
//! goingson carries four escapers and 543 call sites that must pick between
//! them, because `escapeHtml` is built on `textContent` serialization and
//! **`textContent` refuses to encode `"`**. That is what makes it unsound in an
//! attribute, and it is the whole reason the choice exists. Its `escape.js`
//! records the finding as the CHRONIC-XSS seal, and its test suite has a gate
//! keeping the unsafe one off the namespace.
//!
//! [`escape`] here is not built on that, so it encodes the quote along with
//! everything else, which makes one function sound in both sinks. The four-way
//! choice does not move into Rust: it disappears. A caller hands an unescaped
//! value to the output only through [`Markup`], which it has to name.
//!
//! # What the description does not carry
//!
//! One thing: the **current value**, which arrives in [`Filling`]. A webview
//! reads it back out of the DOM, an immediate-mode renderer writes through a
//! `&mut`, and a terminal keeps an edit buffer; a description carrying it would
//! have to carry a way to write it back, at which point it is a form model.

use crate::{Emit, class, status_edge, tone_rules};
use makeover_layout::{Depth, Intent as _, Tone, Value};
use std::fmt::Write as _;

/// Every class a field's markup can carry.
///
/// [`crate::facet::FACET_CLASSES`]' obligation. Every one of these is ruled by
/// the generated sheet, the group's four through `group_rules`; the list is
/// written down anyway, because it is what the hosts' markup is held against
/// and what the vocabulary test holds the sheet against.
///
/// What goes wrong without it: an app checking its stylesheet against
/// [`crate::vocabulary::names`] concludes that its live `.form-group` and
/// `.form-label` rules match nothing and are safe to delete.
pub const FIELD_CLASSES: &[&str] = &[
    "field",
    "form-checkbox-label",
    "form-checklist",
    "form-editor-modes",
    "form-editor-preview",
    "form-error",
    "form-group",
    "form-group-inline",
    "form-hint",
    "form-interval",
    "form-label",
    "form-note",
    "form-option-detail",
    "form-option-reason",
    "form-radio-group",
    "form-radio-label",
    "form-unit",
];

// `form-suggestions`, `form-suggestion` and `form-suggestion-detail` are
// deliberately absent: [`suggestion_rules`] writes their look and
// `quasi-webview` writes their markup, because a suggestion source is a route
// and no description layer carries one. They reach the vocabulary through the
// generated sheet, which is where a name this crate rules but does not emit
// belongs.

/// The state classes a field carries, which take no prefix.
///
/// `chosen` and `latched`'s convention, stated in
/// [`crate::vocabulary::vocabulary`]: a state qualifies a prefixed component
/// (`.mk-form-group.has-error`) rather than standing on its own, so a prefix
/// moves the thing and not its state.
///
/// `has-error` marks the group and `visible` marks the message: a renderer
/// with no descendant selectors cannot find the group from the message, so
/// both are told.
pub const FIELD_STATE_CLASSES: &[&str] = &["has-error", "visible"];

/// A string that is already markup, and is emitted without escaping.
///
/// The one hole in the escaping, and it has to be named to be used. goingson
/// has two live callers that need it, both passing a recurrence-config block
/// built elsewhere, and both would otherwise have their markup rendered as
/// visible angle brackets. A caller constructing this is stating that the
/// contents are trusted; nothing here can check that for them.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct Markup<'a>(pub &'a str);

/// Everything about the field that the description does not carry.
#[derive(Debug, Clone, Copy, Default)]
pub struct Filling<'a> {
    /// What the field holds now.
    pub value: Value<'a>,
    /// Markup appended inside the group, after the hint. Not escaped.
    pub trailing: Option<Markup<'a>>,
    /// Attributes written onto the control element itself. Not escaped.
    ///
    /// [`trailing`](Self::trailing)'s argument at attribute scale: a host knows
    /// facts about the control that no description layer carries. quasi's
    /// suggestion source is the first caller: a field that owns a list of
    /// candidates is a `role="combobox"` pointing at the list it owns, and
    /// neither half is anything the described field can say.
    ///
    /// Written verbatim, so a caller supplies `attr="value"` pairs with no
    /// leading space and does its own escaping. It is [`Markup`]'s hole in the
    /// same wall, named the same way so a caller has to state that the contents
    /// are trusted.
    ///
    /// A radio group drops them, and that is deliberate rather than an
    /// oversight: a radio group is a set of sibling inputs with no one control
    /// element, so there is nowhere honest to put an attribute meant for the
    /// control. The group carries the descriptions for the same reason.
    pub control_attrs: Option<Markup<'a>>,
    /// Scopes the `id` attributes to one instance of the form.
    ///
    /// The field's `name` is what the value submits under and is the same
    /// wherever the form appears; its `id` has to be unique in the document,
    /// and those two facts stop agreeing the moment a form appears twice.
    /// goingson hits this directly: its new-task and edit-task modals are the
    /// same field set, so it prefixes `form-modal-task-new` or `-edit` to keep
    /// `label for` and `aria-describedby` pointing at the right control.
    ///
    /// Applies to `id`, `for` and the `-hint` / `-error` associations. Never to
    /// `name`, which would change what the form submits.
    pub id_prefix: Option<&'a str>,
}

impl<'a> Filling<'a> {
    /// A filling that carries a value and nothing else.
    #[must_use]
    pub const fn of(value: Value<'a>) -> Self {
        Self {
            value,
            trailing: None,
            control_attrs: None,
            id_prefix: None,
        }
    }
}

/// Encode the five characters that let a value stop being a value, into a
/// buffer the caller already has.
///
/// The form an emitter uses. [`escape`] is this with a `String` allocated
/// around it, and the allocation is the whole difference: a described screen
/// escapes once per attribute and once per run of text, so a function that
/// returns a `String` allocates a few thousand times to produce one page,
/// where a template engine writes its escaped bytes straight into the output
/// buffer.
///
/// Sound in element text and in a double-quoted attribute alike, which is the
/// property `textContent`-based escaping cannot have. Both sinks are covered by
/// one function so that no call site has to choose, here or downstream.
///
/// Copies in runs rather than per character. All five encoded characters are
/// ASCII, so a byte scan cannot land inside a multi-byte character and the
/// slice between two of them is always a valid `&str`. Text with nothing to
/// encode, which is most text, is one `push_str` of the whole thing.
pub fn escape_into(text: &str, out: &mut String) {
    let mut start = 0;
    for (index, byte) in text.bytes().enumerate() {
        let encoded = match byte {
            b'&' => "&amp;",
            b'<' => "&lt;",
            b'>' => "&gt;",
            b'"' => "&quot;",
            b'\'' => "&#39;",
            _ => continue,
        };
        out.push_str(&text[start..index]);
        out.push_str(encoded);
        start = index + 1;
    }
    out.push_str(&text[start..]);
}

/// Encode the five characters that let a value stop being a value.
///
/// [`escape_into`] with a buffer of its own, for the callers that want a value
/// rather than an append: a caller assembling an attribute out of several
/// pieces, and everything outside this crate that took this function before the
/// buffer-writing form existed. Emitting into a buffer you already hold is the
/// cheaper path and the one the hosts' emitters take.
#[must_use]
pub fn escape(text: &str) -> String {
    let mut out = String::with_capacity(text.len());
    escape_into(text, &mut out);
    out
}

/// The rules the markdown editor's chrome needs.
///
/// The two names here have no app counterpart to keep, because the chrome did
/// not exist before the member did. The markup is a Write/Preview pair of
/// segments and an empty preview pane, drawn by the hosts; converting markdown
/// stays the host's, where its sanitiser already is.
///
/// Every rule is gated on `[data-format="markdown"]`, which is what keeps them
/// off a plain textarea, and every rule that hides content is gated on
/// `data-ready` as well, which is what keeps them out of a document with no
/// script.
pub(crate) fn editor_rules(opts: &Emit) -> String {
    let mut css = String::new();
    let modes = class("form-editor-modes", opts);
    let preview = class("form-editor-preview", opts);
    let field = class("field", opts);

    // Hidden until something binds the editor: a control rendered into a
    // document with no script looks live and answers nothing, so a reader with
    // no script gets the textarea alone.
    let _ = writeln!(
        css,
        "[data-format=\"markdown\"] > .{modes} {{\n    display: none;\n}}"
    );
    // Block, and nothing about how the two segments sit in it. A button is
    // inline already, so they make a row without this crate saying so, and
    // saying so is where a gap would follow, a magnitude, and
    // `makeover-geometry`'s.
    let _ = writeln!(
        css,
        "[data-format=\"markdown\"][data-ready] > .{modes} {{\n    display: block;\n}}"
    );

    // The pane is empty until the host fills it, so it is out of flow in every
    // state but the one where a bound editor is showing it. An empty box under
    // the control is chrome claiming a preview nobody rendered.
    let _ = writeln!(
        css,
        "[data-format=\"markdown\"] > .{preview} {{\n    display: none;\n}}"
    );
    let _ = writeln!(
        css,
        "[data-format=\"markdown\"][data-ready][data-mode=\"preview\"] > .{preview} \
         {{\n    display: block;\n}}"
    );
    // One at a time. The source and the preview are the same content read two
    // ways, and a field showing both answers its own question twice.
    let _ = writeln!(
        css,
        "[data-format=\"markdown\"][data-ready][data-mode=\"preview\"] > .{field} \
         {{\n    display: none;\n}}"
    );

    // The pane stands where the control stood, so it reads as the surface the
    // control was: `.field` is a well, and this is the well it stands in for.
    // Nothing about size, how tall a preview is is the app's, the way the
    // height of a track is.
    let _ = write!(
        css,
        "[data-format=\"markdown\"] > .{preview} {{\n{}}}\n",
        crate::depth_declarations(Depth::Well)
    );

    css
}

/// The rules a field's group needs: its label, its hint, its message, and the
/// arrangement of the controls that answer one question.
///
/// These were the apps' names from phase A, left unruled so that adoption would
/// delete goingson's `renderFormField` rather than restyle anything. Adoption
/// happened and the argument went with it: a described form in an app that had
/// never written the rules drew its label as body text and its error as a plain
/// sentence, and MNW was that app. wiki `look-restoration`: the renderer owns
/// the look.
///
/// The values are the ones goingson shipped, less the group's own margin. A
/// form spaces its groups with a gap, so a margin here would space them twice;
/// an app laying groups out in normal flow keeps a margin of its own. The radio,
/// checkbox, reason and interval rules came from quasi-webview's arrangement
/// sheet unchanged, where they styled names this crate emits.
///
/// `visible` is the message's state rather than decoration. goingson's scripts
/// raise and lower it on a message that stays in the document, and this crate
/// writes it on every message it emits, so a lowered message is hidden and a
/// written one shows.
///
/// `form-group-inline` is a group sitting in a run: its label goes beside the
/// control rather than over it, so a picker on a line of chips is one line
/// tall.
///
/// `has-error` draws a danger edge down the group's leading side. The control
/// already carries the danger edge through `aria-invalid`; the group's edge is
/// what a reader scanning a long form finds first, and it says which question
/// failed without colouring a word of it. The label and the message keep
/// content ink: a status colour is a fill and an edge, never the colour of
/// words a reader has to read (wiki `quasi-vocabulary-findings`).
pub(crate) fn group_rules(opts: &Emit) -> String {
    let group = class("form-group", opts);
    let label = class("form-label", opts);
    let hint = class("form-hint", opts);
    let error = class("form-error", opts);
    let radios = class("form-radio-group", opts);
    let checklist = class("form-checklist", opts);
    let radio = class("form-radio-label", opts);
    let checkbox = class("form-checkbox-label", opts);
    let reason = class("form-option-reason", opts);
    let interval = class("form-interval", opts);
    let danger = Tone::Danger.token();
    let mut css = String::new();

    let _ = writeln!(
        css,
        ".{label} {{\n    display: block;\n    margin-block-end: var(--gap-bound);\n    color: var(--content);\n    font-weight: bold;\n}}"
    );
    let _ = writeln!(
        css,
        ".{group}.has-error {{\n{}}}",
        status_edge(opts.focus_width, danger)
    );
    // A field in a run: the label beside the box, the way a chip run's word
    // opens it, and in the same weight as that word. Two tracks rather than a
    // wrapping row, because a control fills its box and would wrap under its
    // own label; whatever explains or refuses the answer spans a line under.
    let inline = class("form-group-inline", opts);
    let field = class("field", opts);
    let _ = writeln!(
        css,
        ".{group}.{inline} {{\n    display: grid;\n    grid-template-columns: auto minmax(0, 1fr);\n    align-items: center;\n    column-gap: var(--gap-bound);\n}}\n\
         .{group}.{inline} > .{label} {{\n    margin-block-end: 0;\n    font-weight: inherit;\n}}\n\
         .{group}.{inline} > :not(.{label}, .{field}) {{\n    grid-column: 1 / -1;\n}}"
    );
    let _ = writeln!(
        css,
        ".{hint} {{\n    margin-block-start: var(--gap-bound);\n    max-inline-size: var(--measure-reading, 42rem);\n    color: var(--content-secondary);\n    font-size: var(--text-note);\n}}"
    );
    let _ = writeln!(
        css,
        ".{error} {{\n    margin-block-start: var(--gap-bound);\n    color: var(--content);\n    font-weight: bold;\n}}"
    );
    let _ = writeln!(css, ".{error}:not(.visible) {{\n    display: none;\n}}");

    // A radio group is a stack of labelled choices, and a choice may say why:
    // the box beside its label rather than centred over it. A checklist is the
    // same stack with a box that stays ticked.
    let _ = writeln!(
        css,
        ".{radios},\n.{checklist} {{\n    display: flex;\n    flex-direction: column;\n    gap: var(--gap-peer);\n}}"
    );
    // Two tracks: the box, and everything it labels. A label that wraps
    // wraps in its own track beside the box, where a wrapping flex row left the
    // box alone on the line above its words, and the box sits on the label's
    // first line rather than above it.
    let _ = writeln!(
        css,
        ".{radio},\n.{checkbox} {{\n    display: grid;\n    grid-template-columns: auto 1fr;\n    \
         align-items: baseline;\n    column-gap: var(--gap-bound);\n}}\n\
         .{radio} > :not(input),\n.{checkbox} > :not(input),\n.{reason} {{\n    grid-column: 2;\n    min-width: 0;\n}}"
    );
    // An option's second line, and an unavailable option's reason, sit under
    // its label in the label's track: the second track above holds every child
    // but the box, each on a row of its own.
    let _ = writeln!(
        css,
        ".{interval} {{\n    display: flex;\n    align-items: center;\n    gap: var(--gap-bound);\n}}"
    );
    css
}

/// The rules a field's note needs.
///
/// Nothing emitted a note before the described field could carry one, so
/// there was no app rule to keep.
///
/// The tones are the four a badge carries, drawn as a leading edge the way a
/// toned sentence draws one; the words keep content ink. The bare class is
/// `content` rather than `content-muted`: a note is a consequence the user is
/// meant to read before answering, so muting it by default would be this crate
/// deciding it does not matter.
pub(crate) fn note_rules(opts: &Emit) -> String {
    let note = class("form-note", opts);
    let mut css = String::new();
    let _ = writeln!(css, ".{note} {{\n    color: var(--content);\n}}");
    tone_rules(&mut css, &format!(".{note}"), |tone| {
        status_edge(opts.focus_width, tone)
    });
    css
}

/// The rule a field's unit needs.
///
/// Nothing emitted a unit before the described field could carry one, so
/// there was no app rule to keep.
///
/// One declaration, and it is the whole look. A unit is a fact about the number
/// beside it rather than a second thing to read, so it takes the muted content
/// intent, the same reading `.figure-caption` and `.track-tick` take, and for
/// the same reason.
///
/// And where it sits, which the first version left to the flow and got wrong. A
/// control fills its group's width, so the span after it wrapped onto a line of
/// its own and "min" hung under the box it measures. A group carrying a unit is
/// two columns instead: the control and its unit share the first row's line,
/// and the label, the hint, the note and the message span both. The control's
/// track is `auto`, so it is as wide as the control and the unit reads straight
/// after the number; the spare room goes to the unit's track, where it is
/// empty. The other way round put "min" at the far edge of the pane.
pub(crate) fn unit_rules(opts: &Emit) -> String {
    let unit = class("form-unit", opts);
    let group = class("form-group", opts);
    let mut css = String::new();
    let _ = writeln!(css, ".{unit} {{\n    color: var(--content-muted);\n}}");
    let _ = writeln!(
        css,
        ".{group}:has(> .{unit}) {{\n    display: grid;\n    grid-template-columns: auto minmax(0, 1fr);\n    column-gap: var(--gap-bound);\n    align-items: center;\n}}"
    );
    let _ = writeln!(
        css,
        ".{group}:has(> .{unit}) > :not(.{unit}, input, select, textarea) {{\n    grid-column: 1 / -1;\n}}"
    );
    css
}

/// The rules an option's second line needs.
///
/// [`unit_rules`]' argument: rule what has no app counterpart to keep. An
/// unruled second line renders identically to the label it sits under, which is
/// a worse default than the hand-written markup it replaces.
///
/// Colour only, and muted, which is the same reading `.form-unit` and
/// `.form-suggestion-detail` take: the line orients the label rather than
/// competing with it. Nothing about placement or spacing, for `unit_rules`'
/// reason, a magnitude asserted here belongs to `makeover-geometry`.
pub(crate) fn option_detail_rules(opts: &Emit) -> String {
    let detail = class("form-option-detail", opts);
    let mut css = String::new();
    let _ = writeln!(css, ".{detail} {{\n    color: var(--content-muted);\n}}");
    css
}

/// The rules a field's suggestion list needs.
///
/// [`editor_rules`]' precedent and its argument: these three have no app
/// counterpart to keep because the list did not exist before the member did.
/// The markup is `quasi-webview`'s, a suggestion source is a route, which no
/// description layer carries, and the look is
/// still this crate's, because a renderer inventing how a list of candidates
/// reads is the drift the vocabulary check exists to catch.
///
/// # In flow, and not floating
///
/// An absolutely positioned list needs a positioned ancestor, and the only
/// candidate is `.form-group`, which is the app's class and deliberately
/// unruled here. So the list stands under the control and moves what is below
/// it. An app that wants it over the form positions the group itself, which is
/// one declaration and is the app's call about its own layout.
///
/// `:empty` is what takes it away, so a route that answers with no candidates
/// leaves no box behind. It is a content question rather than a whitespace one
/// only because the emitter writes no whitespace inside the container, which is
/// stated in `quasi-webview`'s own test.
///
/// # Placement, and nothing about size
///
/// One candidate to a line, its detail at the line's end. Without it the
/// entries are bare anchors and run together as one paragraph: MNW's discover
/// search drew "Blue HourProjectKickItem". Spacing comes from the geometry
/// tokens every other form rule uses, so no magnitude is asserted here.
///
/// No height and no scroll ceiling. How tall a list of candidates gets to be
/// before it scrolls is a magnitude, and magnitudes are `makeover-geometry`'s,
/// exactly as the preview pane's height is.
pub(crate) fn suggestion_rules(opts: &Emit) -> String {
    let list = class("form-suggestions", opts);
    let entry = class("form-suggestion", opts);
    let detail = class("form-suggestion-detail", opts);
    let mut css = String::new();

    let _ = writeln!(css, ".{list}:empty {{\n    display: none;\n}}");
    let _ = writeln!(
        css,
        ".{list} {{\n    display: flex;\n    flex-direction: column;\n}}"
    );
    let _ = writeln!(
        css,
        ".{entry} {{\n    display: flex;\n    align-items: baseline;\n    justify-content: space-between;\n    gap: var(--gap-group);\n    padding: var(--gap-bound) var(--gap-peer);\n    color: inherit;\n    text-decoration: none;\n}}"
    );
    // Over what it covers, which is what a list of candidates is even in flow:
    // it is answering the box above it and goes away when the answer is taken.
    css.push_str(&crate::depth_rule(&list, Depth::Overlay));
    // An entry answers a click, so it gets every state one implies.
    css.push_str(&crate::interactive_rules(&entry, Depth::Flat, opts));
    // The keyboard's highlight and the pointer's are the same surface. They are
    // the same fact told two ways, and a list where arrowing and hovering look
    // different is a list that has two current entries.
    //
    // Keyed on `aria-selected` rather than on a class, for the reason
    // `aria-invalid` carries the error state: it is what a screen reader hears,
    // so a look keyed on it cannot drift from what is announced. A `.current`
    // class would also be a name apps already spell for their own reasons:
    // the MNW server has one, and unlayered app CSS beats this layer in
    // silence.
    let _ = writeln!(
        css,
        ".{entry}[aria-selected=\"true\"] {{\n    background: var(--hover-surface);\n}}"
    );
    // The second line, muted rather than disabled. A candidate carries no
    // `unavailable`, and what sits beside the label now is what tells one row
    // from another that reads the same. Disabled would say the row cannot be
    // picked, which is the opposite of what the detail is for.
    let _ = writeln!(css, ".{detail} {{\n    color: var(--content-muted);\n}}");

    css
}

#[cfg(test)]
mod tests;