use crate::{Choice, Curve, ThemeChoice, Tone};

// Names this module's prose links to, resolved for rustdoc.
#[allow(unused_imports)]
use crate::{Awaiting, Contrast, Fill, ThemeVariant};

/// 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.
#[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,
    /// A number inside bounds the user drags across, where the range being
    /// visible is the point.
    ///
    /// Not [`Number`](Self::Number) with [`min`](Field::min) and
    /// [`max`](Field::max), which is the reading to resist and is the same
    /// resistance [`Radio`](Self::Radio) needed against `Select`. A bounded
    /// number and a validated number are different *questions*. A validated
    /// number is typed and can be wrong: the bounds are a rule the answer is
    /// checked against, and being told "must be at least 1" afterwards is the
    /// normal course of it. A range cannot be out of range at all, because the
    /// bounds are the control's extent rather than a rule, and the two ends are
    /// what the question means — audiofiles asks for a classifier threshold
    /// between 0 and 1, where 0 is never and 1 is only-on-certainty, and a typed
    /// 0.72 says nothing without both ends on screen beside it.
    ///
    /// A renderer cannot infer which one is meant from `min`/`max` alone, which
    /// is why this is a kind and not an inference: goingson's `min="1"` duration
    /// is a validated number and would become a slider.
    ///
    /// The membership test passes without stretching: a webview emits
    /// `<input type="range">`, egui has `Slider`, a terminal draws a bar and
    /// takes arrow keys, a CLI takes a bounded argument.
    ///
    /// # It owes its bounds
    ///
    /// [`min`](Field::min) and [`max`](Field::max) are `Option` for every other
    /// kind and are **required** here, in the sense the description can require
    /// anything: [`Field::bounded`] is the check, and a range missing one has no
    /// extent for a renderer to draw. What a renderer does with an unbounded
    /// range is its own call and both answers are honest — fall back to a typed
    /// number, or pick a host default — so this is stated rather than enforced,
    /// the way every other constraint here is.
    ///
    /// [`Field::step`] is the third fact and is genuinely optional: absent, the
    /// host's own granularity stands.
    Range,
    /// One question with two ends: a lower value and an upper one, submitted
    /// under two names.
    ///
    /// "Show me samples between 90 and 130 BPM" has a single answer with two
    /// ends, and the ends constrain each other: a minimum above the maximum is
    /// not a wrong value, it is an empty result nobody asked for. Described as
    /// two [`Number`](Self::Number) fields that is unsayable — nothing says they
    /// are one question, so a renderer draws two controls with two labels and no
    /// relationship, and [`Field::error`] can only be attached to one side of a
    /// fault that belongs to both.
    ///
    /// Not [`Range`](Self::Range), which was the reading to resist and the
    /// resistance is the same one `Range` itself needed against `Number`. A
    /// range describes *one* value inside an extent; this describes two, and the
    /// extent is a bound on each rather than the question's meaning. The two
    /// come apart in the answer: a range has a value, an interval has a pair,
    /// and either end may be absent while the other stands.
    ///
    /// # It states both names
    ///
    /// [`Field::name`] is the lower end and [`Field::upper_name`] is the upper
    /// one, stated rather than derived. One member instead of a naming
    /// convention this crate would then own forever.
    ///
    /// Direction is carried by which member the name sits in, so nothing
    /// separate says which end is which.
    ///
    /// # What it does not enforce
    ///
    /// The crossing rule. A lower end above the upper one is describable here
    /// and always was, exactly as an out-of-[`min`](Field::min) number is: this
    /// crate carries constraints and never checks them, and deciding a value is
    /// wrong stays with whoever validated. What the description buys is that the
    /// fault now has one place to be reported rather than two.
    ///
    /// # Both ends take the same facts
    ///
    /// [`min`](Field::min), [`max`](Field::max), [`step`](Field::step) and
    /// [`unit`](Field::unit) describe the axis rather than one end of it, so
    /// they are read once and applied to both. Six of audiofiles' filter axes
    /// are exactly this: one extent, one unit, one granularity, two ends.
    ///
    /// The bounds are optional here, unlike `Range`. They are a rule the answer
    /// is checked against rather than the control's extent, which is
    /// [`Number`](Self::Number)'s arrangement and not a slider's.
    Interval,
    /// 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.
    Email,
    /// A URL. Same reasoning as [`Email`](Self::Email).
    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.
    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.
    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.
    DateTime,
    /// Several lines of text.
    Textarea,
    /// Several lines of text the user writes markdown in.
    ///
    /// The editing counterpart of prose a description carries as markdown
    /// source, and the reason it can exist at all is the same one that lets the
    /// source be carried: editing markdown is editing text, so a terminal, an
    /// immediate-mode host and a webview all have an honest answer, and none of
    /// them has to refuse. A kind that meant "rich text" in the WYSIWYG sense
    /// would have been a document model, and two of the three hosts would have
    /// had to draw something they cannot.
    ///
    /// What the mark buys over [`Textarea`](Self::Textarea) is that a renderer
    /// may offer the affordances markdown has and plain text does not — a
    /// preview, a syntax pass, a monospaced face for the source — and that a
    /// host reading the value back knows what it is holding. A renderer with
    /// none of that draws a textarea, which is why this is additive rather than
    /// a second control.
    ///
    /// It says nothing about **when** the value is saved. Autosave is a clock,
    /// clocks are not described here, and the four MNW editors this was measured
    /// against each keep their own.
    ///
    /// Sanitising stays where it already is for markdown that is only displayed:
    /// with the renderer, at the point markup is produced. Being described is
    /// not a safety property, and a host with its own sanitiser and its own
    /// content-security posture still owns both.
    Rich,
    /// 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.
    ///
    /// Everything here is 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. No member is ever "the last one".
    Radio,
    /// On or off.
    Checkbox,
    /// A file the user picks from wherever the host keeps files.
    ///
    /// 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.
    ///
    /// # The four things an upload says, and where each of them lives
    ///
    /// | axis | where |
    /// |---|---|
    /// | what it accepts | [`Field::accept`] |
    /// | one file or several | [`Field::multiple`] |
    /// | where the bytes go | the router's action, not here |
    /// | how far along it is | [`Awaiting`] on that action |
    ///
    /// Only the first two are this crate's, and that split is the answer to
    /// "describe an upload in full" rather than a gap in it. A destination is an
    /// address and this crate holds no addresses; progress is a live number and
    /// a description is built once, so the number is the renderer's to observe
    /// against the size [`Awaiting::amount`] carried before the transfer began.
    ///
    /// # How the file is handed over is the host's
    ///
    /// A drop area, a button opening a native picker, a path typed at a prompt:
    /// all three are the same field, and every measured site has the first. It
    /// is not described for the reason no gesture is — this crate owns no
    /// coordinates and no pointer, and a terminal that cannot be dropped on
    /// would be refusing a description it can otherwise honour completely.
    ///
    /// [`Field::accept`] and [`Field::multiple`] are measured rather than
    /// deferred. A member designed against nothing is the rule to keep: count
    /// the sites before adding one.
    File,
    /// Which theme the app wears.
    ///
    /// The one member here that names a *subject* rather than a shape of
    /// answer, and it is worth saying why that is not the door it looks like.
    /// Every other kind is a question a screen might ask about anything; this
    /// one is a specific question every app in the family asks, once, on its
    /// settings screen, and three of them wrote the same control by hand.
    ///
    /// # It is furniture, and the measurement is what says so
    ///
    /// The reading to resist is that this is [`Select`](Self::Select) with a
    /// grouped option list. Max rejected that: `optgroup` appears at one live
    /// site in the tree and the non-theme grouping count is zero, so the thing
    /// that recurs is this picker rather than option lists that group.
    ///
    /// # What it carries that a select cannot
    ///
    /// [`Field::themes`] rather than [`Field::options`], because a theme is
    /// four facts and an option is two. The two extra facts are the ones no
    /// app can supply without redoing work the theme layer has already done:
    /// which [`ThemeVariant`] group a theme is in, and how legible its muted
    /// text measured. `Choice::new(id, format!("{name} ({variant})"))` is what
    /// the three apps had, and it flattens the group into prose and loses the
    /// tier entirely.
    ///
    /// [`Field::follows`] carries the entry that is not a theme.
    ///
    /// # The cost, stated rather than discovered later
    ///
    /// This puts one screen's shape into a vocabulary that otherwise holds
    /// none, which was the objection raised against it and accepted going in.
    /// The mitigation is narrowness: this describes a theme picker, not a
    /// general "list the host resolved" mechanism. A second host-resolved list
    /// is when that generalisation gets measured, and not before.
    ///
    /// A renderer that has not heard of it draws a select over
    /// [`Field::themes`]' names and loses the grouping, which is the state
    /// every app was in before this member. Degrading to the status quo ante
    /// is the floor the member is designed against.
    Theme,
    /// 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)
    }

    /// Whether the kind reads [`Field::themes`] and [`Field::follows`].
    ///
    /// One member answers yes, and it gets a name for
    /// [`takes_files`](Self::takes_files)'s reason rather than in spite of
    /// being alone: four renderers ask it before they read either member, and
    /// a `matches!` per renderer is where the next one goes missing.
    ///
    /// Deliberately not folded into
    /// [`offers_options`](Self::offers_options). A theme picker offers no
    /// [`Choice`]es at all, so a renderer walking `options` for it walks an
    /// empty slice and draws an empty control.
    #[must_use]
    pub const fn offers_themes(self) -> bool {
        matches!(self, Self::Theme)
    }

    /// Whether the value runs to more than one line.
    ///
    /// Named once here for [`temporal`](Self::temporal)'s reason: two kinds
    /// answer yes, every renderer has to ask it before it can size anything,
    /// and a `matches!` per renderer is the pair drifting apart one member at a
    /// time. What a host does with the markdown, if anything, it reads from the
    /// kind itself; this is only whether one line is enough.
    #[must_use]
    pub const fn multiline(self) -> bool {
        matches!(self, Self::Textarea | Self::Rich)
    }

    /// Whether the value is a file the host picks rather than a string typed
    /// into a box.
    ///
    /// One member answers yes, which is [`visible`](Self::visible)'s and
    /// [`confidential`](Self::confidential)'s footing rather than a departure
    /// from it: the question gets a name because three renderers ask it before
    /// they can read [`Field::accept`] or [`Field::multiple`], and a `matches!`
    /// per renderer is where a second file-taking kind would go missing.
    #[must_use]
    pub const fn takes_files(self) -> bool {
        matches!(self, Self::File)
    }

    /// Whether the value is a quantity, so [`Field::unit`] means something.
    ///
    /// The numeric kinds and nothing else. A date is a quantity in the sense
    /// that it is ordered, and it is not one in the sense that matters here:
    /// its unit is fixed by the kind, so `Date` carrying `days` would be the
    /// description restating what [`kind`](Field::kind) already said.
    ///
    /// [`takes_files`](Self::takes_files)'s footing, and for its reason: the
    /// renderers ask this before they decide where a unit goes, and a
    /// `matches!` per renderer is where the next measurable kind goes missing.
    ///
    /// [`Interval`](Self::Interval) is measurable too: an axis is measured in
    /// something and both its ends are in it.
    #[must_use]
    pub const fn measurable(self) -> bool {
        matches!(self, Self::Number | Self::Range | Self::Interval)
    }
}

/// A family of media a file can belong to.
///
/// Three members, because three is what a media type's own first segment offers
/// that a renderer can do anything with. `text` and `application` are families
/// too and neither buys a disclosure — there is no preview of an
/// `application/octet-stream` — so naming them would be a member added for a
/// case nobody has.
///
/// It is the answer to "which disclosure", not a validation rule.
/// [`Field::accept`] is what a host filters on.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub enum Family {
    /// A still picture.
    Image,
    /// Sound.
    Audio,
    /// Moving pictures, with or without sound.
    Video,
}

impl Family {
    /// The wildcard media type that means the whole family.
    ///
    /// `image/*` and its two siblings, which is what the measured sites write
    /// and what a webview puts in an `accept` attribute. Named here so the three
    /// renderers do not each spell the star.
    #[must_use]
    pub const fn wildcard(self) -> &'static str {
        match self {
            Self::Image => "image/*",
            Self::Audio => "audio/*",
            Self::Video => "video/*",
        }
    }

    /// The family a media type's first segment names, if it is one of these.
    ///
    /// Case-insensitive on the segment, because a media type is
    /// case-insensitive and half the tree writes them lowercase by habit rather
    /// than by rule.
    #[must_use]
    pub fn of_type(media_type: &str) -> Option<Self> {
        let (top, _) = media_type.split_once('/')?;
        if top.eq_ignore_ascii_case("image") {
            Some(Self::Image)
        } else if top.eq_ignore_ascii_case("audio") {
            Some(Self::Audio)
        } else if top.eq_ignore_ascii_case("video") {
            Some(Self::Video)
        } else {
            None
        }
    }
}

/// One entry in a file field's accept list.
///
/// Three shapes rather than a string, and all three are in the measured sites:
/// the MNW server writes `image/*`, `image/jpeg,image/png,image/webp`,
/// `.zip,.dmg,.exe,.appimage,.deb,.tar.gz,.clap,.vst3` and, in one place,
/// `.csv,text/csv`. A single string would carry all of them and answer nothing
/// about any of them.
///
/// # Why the list is not just a filter
///
/// It is read twice. Once to decide what the picker offers, which any of the
/// three shapes serves, and once to decide **which disclosure** the field gets:
/// a preview for a picture, a duration or a waveform for a sound. There is one
/// upload shape and a media upload is that shape with more of it shown, so the
/// accept list is what says which more. [`family`](Self::family) is that
/// question answered once here instead of a media-type parser in each renderer.
///
/// # A suffix names no family, on purpose
///
/// `.mp3` is audio in fact, and nothing here says so. A suffix-to-family table
/// in a published crate is a mapping that goes stale, disagrees with the host's
/// own idea of what a file is, and is wrong the first time somebody hands it a
/// container. A call site that wants a picture's preview writes
/// [`Family::Image`] or `image/jpeg`; a call site listing installer suffixes
/// wants no disclosure anyway, which is the measured case.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub enum Accepted<'a> {
    /// Every file of a family: `image/*` and its siblings.
    Family(Family),
    /// One media type, written the way a media type is written:
    /// `image/jpeg`, `text/csv`.
    Type(&'a str),
    /// One file-name suffix, written with its leading dot: `.zip`, `.tar.gz`.
    ///
    /// A suffix and not an extension, because `.tar.gz` is a measured site and
    /// is two dots.
    Suffix(&'a str),
}

impl<'a> Accepted<'a> {
    /// The family this entry belongs to, when it names one.
    ///
    /// [`None`] for a [`Suffix`](Self::Suffix) and for any media type outside
    /// the three families, which is the honest answer rather than a missing
    /// one: the description did not say.
    #[must_use]
    pub fn family(self) -> Option<Family> {
        match self {
            Self::Family(family) => Some(family),
            Self::Type(media_type) => Family::of_type(media_type),
            Self::Suffix(_) => None,
        }
    }

    /// How a host that wants one string writes this entry.
    ///
    /// A webview's `accept` attribute takes exactly these spellings, and a
    /// terminal listing what it will take reads the same words.
    #[must_use]
    pub const fn as_str(self) -> &'a str {
        match self {
            Self::Family(family) => family.wildcard(),
            Self::Type(text) | Self::Suffix(text) => text,
        }
    }
}

#[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.
    ///
    /// The *lower* end's name for a [`FieldKind::Interval`], whose upper end is
    /// [`upper_name`](Self::upper_name). Every other kind submits one value and
    /// this is the whole of it.
    pub name: &'a str,
    /// The name a [`FieldKind::Interval`]'s upper end is submitted under.
    ///
    /// [`None`] for every other kind, and sayable-and-ignored there the way
    /// [`options`](Self::options) is on a kind that offers none.
    ///
    /// Stated rather than derived from [`name`](Self::name), and
    /// [`FieldKind::Interval`] carries the measurement that decided it: the two
    /// sites in this tree disagree about affix order, so a derived rule would
    /// rename one of them. Which member a name sits in is also what says which
    /// end it is, so nothing separate carries the direction.
    ///
    /// An interval missing it is an interval with one end that can be submitted,
    /// which is a description a renderer may draw honestly and no better than
    /// that. [`Field::interval`] is what makes forgetting it unsayable, on the
    /// same footing as [`Field::range`] and its bounds.
    pub upper_name: Option<&'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>,
    /// A consequence of the answer the user has given, carrying its own tone.
    ///
    /// The third message channel, between [`hint`](Self::hint) and
    /// [`error`](Self::error) and overlapping neither. A hint is standing help
    /// that does not depend on the value; an error says the value is not
    /// acceptable. A note is the case in the middle: the value is perfectly
    /// acceptable and choosing it costs something the user should know about.
    ///
    /// The first consumer is audiofiles' export Format field, where choosing
    /// WAV or AIFF over Original re-encodes and silently drops embedded BWF,
    /// iXML, loop points, cue markers and ID3. That is not a validation
    /// failure and it is not standing help — it is true of one answer to one
    /// question — and it was hand-drawn in the app's own draw callback for
    /// want of anywhere to say it.
    ///
    /// The tone is carried rather than fixed at [`Tone::Warning`] because the
    /// channel is not only for warnings: the same slot says "this is the
    /// recommended one" ([`Tone::Success`]) and "this is what that setting
    /// implies" ([`Tone::Info`]). A renderer gets the announcement behaviour
    /// off the tone for free — makeover-webview emits `data-tone` and treats
    /// Warning and Danger as assertive for `aria-live`.
    ///
    /// It does **not** make the field invalid. [`invalid`](Self::invalid) stays
    /// `error.is_some()`, so a note never marks the group as a problem.
    ///
    /// # Precedence, for a renderer with room for one
    ///
    /// Error, then note, then hint. A renderer that shows every message shows
    /// them in that order too. makeover-tui is the one with room for exactly
    /// one line, and it is why the order is decided here rather than three
    /// times: what is wrong outranks what it costs, which outranks how it
    /// works.
    pub note: Option<(Tone, &'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.
    ///
    /// 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>],
    /// The themes offered, in the order they are offered.
    ///
    /// Empty for every kind [`FieldKind::offers_themes`] rejects, and sayable
    /// as empty for the one that accepts it: an app whose theme directories
    /// hold nothing has a picker offering only [`follows`](Self::follows),
    /// which is a true description of that machine.
    ///
    /// **The order is the grouping.** Entries arrive sorted by
    /// [`ThemeVariant`] and then by [`Contrast`] within each variant, so a
    /// renderer that draws headings walks the run of one variant and a renderer
    /// that cannot still gets the useful order. Handing back groups would force
    /// the second renderer to flatten what the first wanted.
    ///
    /// Nothing here sorts. The description carries the order it was given, and
    /// the sort belongs with whoever measured the tiers — `makeover::theme_options`
    /// is what produces it, and re-sorting here would be this crate deciding a
    /// question it cannot see the inputs to.
    ///
    /// Which theme is *current* is not here. That is the value, and the value
    /// is renderer state, exactly as it is for [`options`](Self::options).
    pub themes: &'a [ThemeChoice<'a>],
    /// The entry that follows the ambient mode instead of naming a theme.
    ///
    /// [`None`] for a picker that does not offer one, which is a real answer:
    /// an app whose host has no ambient mode to follow should not offer a row
    /// that does nothing.
    ///
    /// A [`Choice`] rather than a bare label, because the *value* is the app's.
    /// Every store in the family spells it `system` today and none of them is
    /// obliged to; a description that hardcoded the spelling would be this
    /// crate holding a fact about somebody else's config table.
    ///
    /// It is not a [`ThemeChoice`] with an absent variant. Following is a
    /// standing instruction that resolves differently as the desktop flips, and
    /// a theme id is an answer that does not — which is the distinction
    /// `makeover::ThemeSelection` exists to hold, carried here rather than
    /// blurred.
    pub follows: Option<Choice<'a>>,
    /// What a file field takes, in the order a host offering the list shows it.
    ///
    /// Empty for every kind [`FieldKind::takes_files`] rejects, and empty is
    /// also a real answer for one that accepts it: a field that takes any file
    /// says so by listing nothing, which is what an `<input type="file">` with
    /// no `accept` does and what most of the measured sites are.
    ///
    /// It is a filter and it is the disclosure cue, and [`Accepted`]'s doc
    /// carries which reading is which. Nothing here validates: a host may hand
    /// back a file the list does not cover, exactly as a browser does when the
    /// user switches the picker to "All Files", and deciding a value is wrong
    /// stays with whoever validated.
    pub accept: &'a [Accepted<'a>],
    /// Whether more than one file may be picked at once.
    ///
    /// Only [`FieldKind::takes_files`] reads it. A multi-valued answer to any
    /// other question is a different shape — a set of options, a repeated
    /// group — and neither is this flag with a different kind beside it.
    ///
    /// False is the common case: 4 of the MNW server's 16 file inputs carry it.
    pub multiple: bool,
    /// Whether the form refuses to submit without it.
    pub required: bool,
    /// The longest the value may be, in characters.
    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>,
    /// The granularity the value moves in, as the host would write it.
    ///
    /// Text for [`min`](Self::min)'s reason, and it earns it twice over: the
    /// step of a date is a day and the step of a threshold is 0.01, and a
    /// numeric member could say one of them.
    ///
    /// Absent means the host's own granularity, which is the honest default
    /// rather than a missing value: a webview's `<input>` steps by 1 unless told
    /// otherwise, and that is the browser's rule and not this crate's to
    /// restate.
    ///
    /// # It is the granularity of a *typed* value
    ///
    /// [`FieldKind::Range`] reads its own from [`curve`](Self::curve) and
    /// ignores this. On a slider the granularity and the mapping are one
    /// decision, and on a typed number there is no mapping to decide with. See
    /// [`Curve`], "Why the step is here".
    pub step: Option<&'a str>,
    /// How a slider's position becomes its value, and how finely it moves.
    ///
    /// [`FieldKind::Range`]'s, and nothing else reads it: a typed number has a
    /// granularity but no mapping, and takes [`step`](Self::step) instead.
    ///
    /// Defaults to [`Curve::Linear`] with no step, which is what an
    /// undescribed range means.
    pub curve: Curve<'a>,
    /// What the number is measured in: `s`, `ms`, `dB`, `GiB`.
    ///
    /// A fact about the value, not part of the question's name, and that
    /// distinction is the whole reason it is a member. The two readings come
    /// apart the moment anything reads a field back rather than drawing it: a
    /// [`max`](Self::max) of `-96` and a bound of `-96 dBFS` are the same number
    /// and not the same answer, and under the convention this replaces the unit
    /// could only be recovered by parsing it back out of a label.
    ///
    /// # Where a renderer draws it
    ///
    /// Beside the value, wherever that host puts a value. Not in the label: a
    /// label is the sentence above the control, so unit-in-label reads the same
    /// on every host and is wrong on any host with somewhere better. egui puts
    /// it inside
    /// the slider where the readout already is, a terminal appends it to the
    /// value in the edit line, a webview sets it adjacent to the input.
    ///
    /// # Which kinds read it
    ///
    /// [`FieldKind::measurable`] answers, and it is
    /// [`takes_files`](FieldKind::takes_files)'s footing: three renderers ask
    /// before they can decide whether to draw this, and a `matches!` per
    /// renderer is where the next measurable kind goes missing. A unit on a kind
    /// that rejects it is sayable and ignored, the same way
    /// [`options`](Self::options) is on a kind that offers none.
    ///
    /// # Why a string
    ///
    /// The measured sites are `GiB`, `dBFS`, `s` and `ms`. An enum would have to
    /// grow a member for every unit any consumer ever wants, and this crate does
    /// not know them; it knows that a number has one.
    ///
    /// Written as the symbol alone, with no brackets and no leading space. The
    /// spacing is the renderer's, because a slider's readout and a sentence want
    /// different answers.
    pub unit: Option<&'a str>,
    /// Whether the field lives behind a "more options" disclosure.
    pub extended: bool,
    /// Whether this local wall-clock value is submitted as an absolute instant.
    ///
    /// [`FieldKind::DateTime`] asks for a time the way a person says one --
    /// "the 14th at half past two" -- and that names a different moment in
    /// Denver than it does in Berlin. A route that stores an instant needs the
    /// moment, so somebody has to convert. This member says the description
    /// wants that conversion; it does not say how.
    ///
    /// # The conversion belongs to the renderer
    ///
    /// Because the renderer is the only party that knows what "your computer's
    /// time zone" means for its host. A browser has one and the user is sitting
    /// in it; a TUI reads the host clock; an egui app reads the same clock a
    /// different way. Nothing above the renderer can answer it, and the
    /// alternatives all try: a hidden IANA-zone field needs a host capability
    /// for reading the zone that three hosts answer differently, plus a kind
    /// that does not exist, plus a wire-contract change; a timezone on the
    /// user's profile is a product decision wearing a bug's clothes. Say it
    /// here, and the next reader does not propose them again.
    ///
    /// # What a renderer does
    ///
    /// Draws the same control it always did -- the flag changes what is
    /// *submitted*, not what is shown -- and converts the local value to an
    /// absolute instant on the way out. A renderer that cannot convert submits
    /// the local value unchanged, which is what every renderer did before this
    /// existed.
    ///
    /// No wire contract moves when a site adopts it: the route was already
    /// receiving an instant. What changes is who computed it.
    ///
    /// # Which kinds read it
    ///
    /// [`FieldKind::DateTime`]'s. `Date` and `Time` are each half a moment and
    /// cannot name one on their own, so the flag is sayable and ignored there,
    /// the way [`options`](Self::options) is on a kind that offers none.
    pub as_instant: 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,
            upper_name: None,
            label,
            hint: None,
            error: None,
            note: None,
            placeholder: None,
            options: &[],
            themes: &[],
            follows: None,
            accept: &[],
            multiple: false,
            required: false,
            max_length: None,
            min: None,
            max: None,
            step: None,
            curve: Curve::Linear { step: None },
            unit: None,
            extended: false,
            as_instant: false,
        }
    }

    /// A bounded number the user drags across its whole extent.
    ///
    /// The third under-described kind, and it gets a constructor for
    /// [`select`](Self::select)'s reason: a range is the one kind whose bounds
    /// are not a rule but the control itself, so a call site that forgot them
    /// has a slider with nothing to slide across. Taking them as arguments is
    /// what makes that unsayable.
    ///
    /// The granularity stays a field rather than a fourth argument, and it is
    /// [`curve`](Self::curve)'s: it is genuinely optional, since the host's own
    /// is a real answer, and the two bounds are not.
    #[must_use]
    pub const fn range(name: &'a str, label: &'a str, min: &'a str, max: &'a str) -> Self {
        Self {
            min: Some(min),
            max: Some(max),
            ..Self::new(FieldKind::Range, name, label)
        }
    }

    /// One question with two ends, taking the name each end submits under.
    ///
    /// A constructor for [`range`](Self::range)'s reason inverted: a range's
    /// bounds are what a call site cannot forget, and an interval's second name
    /// is. An interval built through [`new`](Self::new) has an upper end with
    /// nowhere to be submitted, and nothing downstream can invent one, so taking
    /// it as an argument is what makes that unsayable.
    ///
    /// The extent, the granularity and the unit stay members. They describe the
    /// axis rather than either end and they are genuinely optional, which is
    /// [`FieldKind::Number`]'s arrangement and the one an interval takes.
    #[must_use]
    pub const fn interval(name: &'a str, upper_name: &'a str, label: &'a str) -> Self {
        Self {
            upper_name: Some(upper_name),
            ..Self::new(FieldKind::Interval, name, label)
        }
    }

    /// A file field, taking the given accept list.
    ///
    /// The fourth under-described kind and it gets a constructor for
    /// [`range`](Self::range)'s reason rather than [`select`](Self::select)'s:
    /// a file field with no accept list is not broken, it is a field that takes
    /// anything, and the hazard is the opposite one. A call site that meant to
    /// restrict and forgot has a picker offering every file on the machine and
    /// a server refusing the upload afterwards, which is the failure the list
    /// exists to move forward. Taking it as an argument is what makes an
    /// accidental omission a deliberate `&[]`.
    ///
    /// [`multiple`](Self::multiple) stays a field. One file is the common case
    /// and the honest default; several is the thing worth saying.
    #[must_use]
    pub const fn upload(name: &'a str, label: &'a str, accept: &'a [Accepted<'a>]) -> Self {
        Self {
            accept,
            ..Self::new(FieldKind::File, name, label)
        }
    }

    /// 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)
    }

    /// A theme picker over the themes the host resolved.
    ///
    /// A constructor for [`select`](Self::select)'s reason and one of its own.
    /// The shared reason: a theme picker built through [`new`](Self::new) has
    /// an empty [`themes`](Self::themes) list and draws an empty control. Its
    /// own: the list is the *only* thing this kind takes that a call site
    /// cannot get wrong by omission and can get wrong by substitution, since
    /// [`options`](Self::options) is right there and reads as if it would work.
    ///
    /// [`following`](Self::following) is the builder rather than a fourth
    /// argument, because a picker with no follow-the-system row is a real
    /// picker and every renderer draws it honestly.
    #[must_use]
    pub const fn theme(name: &'a str, label: &'a str, themes: &'a [ThemeChoice<'a>]) -> Self {
        Self {
            themes,
            ..Self::new(FieldKind::Theme, name, label)
        }
    }

    /// The same picker, offering a row that tracks the ambient mode.
    ///
    /// The [`Choice`] carries the value the app's own store spells it with.
    #[must_use]
    pub const fn following(mut self, follow: Choice<'a>) -> Self {
        self.follows = Some(follow);
        self
    }

    /// 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.
    ///
    /// [`note`](Self::note) is deliberately not consulted. A note says the
    /// answer costs something, not that it is unacceptable, and a field the
    /// user may submit as it stands is not invalid.
    #[must_use]
    pub const fn invalid(&self) -> bool {
        self.error.is_some()
    }

    /// Whether the field carries both ends of its extent.
    ///
    /// Only [`FieldKind::Range`] owes them, and it owes them absolutely: a
    /// slider with one end missing has no extent to draw. Named here rather
    /// than left to each renderer to test `min.is_some() && max.is_some()`,
    /// which is three renderers arriving at the same condition and one of them
    /// getting it wrong, and named as a question about the *field* rather than
    /// about the kind because the kind cannot see the bounds.
    ///
    /// It is a check and not a guarantee. Nothing here refuses to build an
    /// unbounded range — [`Field::range`] is what makes the bounded one easy —
    /// so a renderer asks this and falls back to whatever its host does
    /// honestly with a number.
    #[must_use]
    pub const fn bounded(&self) -> bool {
        self.min.is_some() && self.max.is_some()
    }

    /// Whether anything in [`accept`](Self::accept) names a media family.
    ///
    /// The question a renderer asks before it decides to keep room for a
    /// preview, and it is deliberately the *whole list* rather than one entry:
    /// the media dropzone this was measured against takes `image/*,video/*`, so
    /// there is no single family to return and there is still a disclosure to
    /// offer. Which one it turns out to be is known once a file is picked, which
    /// is renderer-side and after the description is gone.
    ///
    /// False for an empty list, for a list of suffixes, and for `text/csv`. A
    /// renderer that wants the family of a particular entry reads
    /// [`Accepted::family`].
    #[must_use]
    pub fn accepts_media(&self) -> bool {
        self.accept.iter().any(|one| one.family().is_some())
    }
}