razeri-plugin-contract 0.16.0

// Copyright (C) 2026 Gabriel Shea
// Licensed under either of the Apache License, Version 2.0 or the MIT
// license, at your option.

//! **How a source writes an amount** — the picture a design states, read once
//! and handed to every engine as a small set of facts rather than as text.
//!
//! A file writes a price `$2,899.99 `. The dollar sign is not part of the
//! number, the comma is not a decimal point, and the trailing space is not
//! anything at all. Handed to an engine as it stands, that text converts to
//! nothing on one engine, to `2` on another and to an error on a third. So the
//! design says how the source writes it, and this is what that statement is.
//!
//! ## The picture
//!
//! A format is written the way a spreadsheet writes one: the value with its
//! digits replaced by digit places.
//!
//! ```text
//! $#,##0.00      $2,899.99        a dollar sign, thousands in groups of
//!                                 three, two places after the point
//! #,##0          2,899            a whole number written in groups
//! 0.000          1.158            three places after the point and no
//!                                 grouping
//! #,##0.00 kr    2 899,00 kr      no — see below
//! ```
//!
//! - `0` and `#` are digit places. Both mean *digits go here*; a reader does
//!   not care which, and the two are allowed side by side because that is how
//!   the picture is normally written.
//! - `,` is the thousands mark and `.` is the decimal point. They mean that
//!   whichever characters the picture uses, so a picture writing them the
//!   other way round is refused rather than read backwards. A source that
//!   writes `2.899,99` is not covered today, and the refusal says so.
//! - Everything before the first digit place is the symbol that opens the
//!   value, and everything after the last digit place is the symbol that
//!   closes it. Either may be empty.
//! - A space around the whole value is always allowed and never written in
//!   the picture. `$2,899.99 ` and `$2,899.99` are the same value.
//! - A minus sign in front of the value is always allowed. A source that
//!   writes a negative in brackets instead says so with a second picture
//!   after a semicolon: `$#,##0.00;($#,##0.00)`.
//!
//! ## What an engine does with it
//!
//! The picture names the marks that are not part of the number. A reader takes
//! those marks out and converts what is left, so the work every engine has to
//! do is the same work, and the value that lands is the same value. What
//! differs between engines is only how each one spells *take this out* and
//! *convert this, and answer nothing if you cannot* —
//! [`read_with_safe_cast`], [`read_with_pattern_guard`] and
//! [`read_with_digit_check`] are those three answers, written once here so an
//! engine states its own vocabulary rather than its own rules.

use std::fmt;

use crate::DataType;
use crate::sid::normalize_trim;
use crate::sid_spelling::{Concat, SidSpelling};

/// How the source writes a value below zero.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Negatives {
    /// A minus sign in front of the value — `-2,899.99`. The ordinary case,
    /// and what a picture with no second half means.
    Minus,
    /// Brackets around the whole value — `($2,899.99)`, the way a ledger
    /// writes one. Stated with a second picture after a semicolon.
    Parentheses,
}

/// What is wrong with a picture, in the words the refusal uses.
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum NumberFormatError {
    /// A picture saying nothing at all.
    Empty,
    /// A picture with no digit place in it.
    NoDigits,
    /// A thousands mark after the decimal point.
    MarkAfterPoint,
    /// More than one decimal point.
    TwoPoints,
    /// A thousands mark with something other than three digit places after it.
    GroupIsNotThree(usize),
    /// A digit place after the symbol that closes the value.
    DigitsAfterSymbol,
    /// More than two pictures separated by semicolons.
    TooManyPictures,
    /// A second picture that is not the first one in brackets.
    NegativeIsNotBracketed,
}

impl fmt::Display for NumberFormatError {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            Self::Empty => f.write_str(
                "the format is empty. Write how the source writes the value, with its digits \
                 replaced by digit places — `$#,##0.00` for `$2,899.99`, `0.000` for `1.158`",
            ),
            Self::NoDigits => f.write_str(
                "the format holds no digit place, so there is no number in it. A digit place is \
                 `0` or `#` — `$#,##0.00` for `$2,899.99`",
            ),
            Self::MarkAfterPoint => f.write_str(
                "the format puts a `,` after the `.`. In a picture `,` is the thousands mark and \
                 `.` is the decimal point, whichever characters the source itself uses, so the \
                 thousands come first. A source that writes `2.899,99` is not covered today",
            ),
            Self::TwoPoints => f.write_str(
                "the format has two decimal points, and a number has one. If one of them is the \
                 thousands mark, write it as `,` — `$#,##0.00` is two thousand eight hundred and \
                 ninety-nine point nine nine",
            ),
            Self::GroupIsNotThree(places) => write!(
                f,
                "the format puts {places} digit {} after a `,`, and a thousands mark \
                 separates groups of three. Write `#,##0` for a number written `2,899`",
                if *places == 1 { "place" } else { "places" }
            ),
            Self::DigitsAfterSymbol => f.write_str(
                "the format has digit places on both sides of the symbol. The symbol opens the \
                 value or closes it — `$#,##0.00` or `#,##0.00 kr` — and a number written in two \
                 halves is not covered",
            ),
            Self::TooManyPictures => f.write_str(
                "the format holds more than two pictures. One picture says how the source writes \
                 a value; a second, after a `;`, says how it writes a negative one",
            ),
            Self::NegativeIsNotBracketed => f.write_str(
                "the second picture is not the first one in brackets. It says the source writes a \
                 negative the way a ledger does, so it is the first picture inside `(` and `)` — \
                 `$#,##0.00;($#,##0.00)`. A source that writes a leading minus needs no second \
                 picture at all",
            ),
        }
    }
}

impl std::error::Error for NumberFormatError {}

/// **How a source writes an amount, read once.**
///
/// Built from the authored picture and never rebuilt: every engine is handed
/// the same facts, so two engines cannot disagree about what the design said.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct NumberFormat {
    spelled: String,
    opens_with: String,
    closes_with: String,
    grouped: bool,
    decimals: usize,
    negatives: Negatives,
}

/// The characters a picture reserves for itself.
const DIGIT_PLACES: [char; 2] = ['#', '0'];
const THOUSANDS_MARK: char = ',';
const DECIMAL_POINT: char = '.';

impl NumberFormat {
    /// Read a picture, or say what is wrong with it.
    ///
    /// # Errors
    ///
    /// Nothing at all, no digit place, a thousands mark in the wrong place or
    /// with the wrong number of places after it, two decimal points, digits on
    /// both sides of the symbol, or a second picture that is not the first one
    /// in brackets.
    pub fn parse(spelled: &str) -> Result<Self, NumberFormatError> {
        if spelled.is_empty() {
            return Err(NumberFormatError::Empty);
        }
        let mut pictures = spelled.split(';');
        let positive = pictures.next().unwrap_or_default();
        let negative = pictures.next();
        if pictures.next().is_some() {
            return Err(NumberFormatError::TooManyPictures);
        }
        let mut read = Self::read_one(positive)?;
        if let Some(negative) = negative {
            if negative != format!("({positive})") {
                return Err(NumberFormatError::NegativeIsNotBracketed);
            }
            read.negatives = Negatives::Parentheses;
        }
        spelled.clone_into(&mut read.spelled);
        Ok(read)
    }

    /// One picture — the symbol that opens the value, the number, the symbol
    /// that closes it.
    fn read_one(picture: &str) -> Result<Self, NumberFormatError> {
        let is_body = |c: char| DIGIT_PLACES.contains(&c) || c == THOUSANDS_MARK;
        let first =
            picture.find(|c| DIGIT_PLACES.contains(&c)).ok_or(NumberFormatError::NoDigits)?;
        let opens_with = picture[..first].to_owned();
        let rest = &picture[first..];

        // The number runs to the end of the digit places, the thousands marks
        // and the one decimal point with its places.
        let mut body = String::new();
        let mut seen_point = false;
        let mut closes_at = rest.len();
        for (at, character) in rest.char_indices() {
            if is_body(character) {
                if seen_point && character == THOUSANDS_MARK {
                    return Err(NumberFormatError::MarkAfterPoint);
                }
                body.push(character);
                continue;
            }
            if character == DECIMAL_POINT {
                if seen_point {
                    return Err(NumberFormatError::TwoPoints);
                }
                seen_point = true;
                body.push(character);
                continue;
            }
            closes_at = at;
            break;
        }
        let closes_with = rest[closes_at..].to_owned();
        if closes_with.contains(DIGIT_PLACES) {
            return Err(NumberFormatError::DigitsAfterSymbol);
        }

        let (whole, fraction) = match body.split_once(DECIMAL_POINT) {
            Some((whole, fraction)) => (whole, fraction),
            None => (body.as_str(), ""),
        };
        if fraction.contains(THOUSANDS_MARK) {
            return Err(NumberFormatError::MarkAfterPoint);
        }
        let grouped = whole.contains(THOUSANDS_MARK);
        if grouped {
            for group in whole.split(THOUSANDS_MARK).skip(1) {
                if group.chars().count() != 3 {
                    return Err(NumberFormatError::GroupIsNotThree(group.chars().count()));
                }
            }
        }

        Ok(Self {
            spelled: picture.to_owned(),
            opens_with,
            closes_with,
            grouped,
            decimals: fraction.chars().count(),
            negatives: Negatives::Minus,
        })
    }

    /// The picture as the design wrote it — what a refusal quotes back.
    #[must_use]
    pub fn spelled(&self) -> &str {
        &self.spelled
    }

    /// The symbol in front of the value, empty where there is none.
    #[must_use]
    pub fn opens_with(&self) -> &str {
        &self.opens_with
    }

    /// The symbol after the value, empty where there is none.
    #[must_use]
    pub fn closes_with(&self) -> &str {
        &self.closes_with
    }

    /// Whether the source writes the thousands in groups.
    #[must_use]
    pub fn grouped(&self) -> bool {
        self.grouped
    }

    /// How many places the source writes after the decimal point.
    #[must_use]
    pub fn decimals(&self) -> usize {
        self.decimals
    }

    /// How the source writes a value below zero.
    #[must_use]
    pub fn negatives(&self) -> Negatives {
        self.negatives
    }

    /// **The marks that are not part of the number**, in the order a reader
    /// takes them out.
    ///
    /// The symbol that opens the value, the symbol that closes it, the
    /// thousands mark, and the brackets where the source writes a negative in
    /// them. What is left is digits, at most one decimal point and at most a
    /// leading minus.
    #[must_use]
    pub fn marks(&self) -> Vec<String> {
        let mut marks = Vec::new();
        if !self.opens_with.is_empty() {
            marks.push(self.opens_with.clone());
        }
        if !self.closes_with.is_empty() {
            marks.push(self.closes_with.clone());
        }
        if self.grouped {
            marks.push(THOUSANDS_MARK.to_string());
        }
        if self.negatives == Negatives::Parentheses {
            marks.push("(".to_owned());
            marks.push(")".to_owned());
        }
        marks
    }

    /// **The shape a value has to have**, as a regular expression, for an
    /// engine whose own conversion raises rather than answering nothing.
    ///
    /// It is built from the same picture the marks come from, so the guard and
    /// the reading cannot hold two opinions about one spelling. The value is
    /// matched with its surrounding spaces already gone.
    #[must_use]
    pub fn shape_pattern(&self) -> String {
        let digits = if self.grouped { "[0-9][0-9,]*" } else { "[0-9]+" };
        let fraction = if self.decimals > 0 { "([.][0-9]+)?" } else { "" };
        let (open, close) = match self.negatives {
            Negatives::Minus => ("-?", ""),
            Negatives::Parentheses => ("[(]?", "[)]?"),
        };
        format!(
            "^{open}{}{digits}{fraction}{}{close}$",
            optional_literal(&self.opens_with),
            optional_literal(&self.closes_with),
        )
    }
}

impl fmt::Display for NumberFormat {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.write_str(&self.spelled)
    }
}

/// One literal run, made safe for a regular expression and made optional.
///
/// Optional because a file that writes `$6.62` on most lines and `6.62` on one
/// has still written a price on both, and a reader that demanded the symbol
/// would set aside a row nothing is wrong with.
fn optional_literal(text: &str) -> String {
    if text.is_empty() {
        return String::new();
    }
    let mut pattern = String::new();
    for character in text.chars() {
        if character.is_ascii_alphanumeric() {
            pattern.push(character);
        } else {
            pattern.push('[');
            pattern.push(character);
            pattern.push(']');
        }
    }
    format!("({pattern})?")
}

/// **Whether this is a type an amount can land in.**
///
/// A picture says how a number is written down. On a name or a day there is
/// nothing for it to be about, so an engine asked to read one into anything
/// else states nothing and razeri refuses the column by name.
#[must_use]
pub fn is_a_number(target: &DataType) -> bool {
    matches!(target, DataType::Integer(_) | DataType::Decimal(_) | DataType::Float(_))
}

/// **The value with the picture's marks taken out**, in this engine's own
/// spelling of `TRIM` and `REPLACE`.
///
/// Those two words are already written down per engine, on the vocabulary the
/// surrogate key is derived through, and a second table of the same two words
/// would be a second answer to one question.
///
/// Parameters:
/// - `format`: the picture the design stated;
/// - `spelling`: this engine's vocabulary;
/// - `expr`: an already-rendered text expression.
#[must_use]
pub fn strip_marks(format: &NumberFormat, spelling: SidSpelling, expr: &str) -> String {
    let replace = spelling.lower.replace;
    let prefix = spelling.lower.pattern_literal_prefix;
    format.marks().iter().fold(normalize_trim(spelling, expr), |text, mark| {
        format!("{replace}({text}, {prefix}'{}', {prefix}'')", mark.replace('\'', "''"))
    })
}

/// **The one thing the picture's marks cannot come out of a value without.**
///
/// Taking the thousands mark out of `2.899,99` leaves `2.89999`, which
/// converts on every engine to a price nobody wrote — the worst answer
/// available, because it is a number and it is wrong. The picture says the
/// thousands come before the decimal point, so a value writing them after it
/// is not written that way, and this is that sentence as a condition.
///
/// `LIKE` and nothing else, because this has to hold on the engines with no
/// way to match a pattern as much as on the ones that have. `None` where the
/// picture names no thousands mark: there is nothing to take out of the wrong
/// place.
fn separators_in_order(format: &NumberFormat, spelling: SidSpelling, expr: &str) -> Option<String> {
    if !format.grouped() {
        return None;
    }
    let prefix = spelling.lower.pattern_literal_prefix;
    Some(format!(
        "{} NOT LIKE {prefix}'%{DECIMAL_POINT}%{THOUSANDS_MARK}%'",
        normalize_trim(spelling, expr)
    ))
}

/// **How an engine asks whether a text holds a character of a class**, for
/// [`read_with_safe_cast`] and [`read_with_digit_check`].
///
/// The readers hold a value to the picture's shape. Flat, each rule is one
/// test beside the next, each naming the value once and nesting nothing, so a
/// value is no deeper than its marks' own `REPLACE`s: an engine whose parser
/// holds a hundred levels — `SQLite` 3.45's — reads a wide table. Two of the
/// rules need a class of characters (a digit; anything but a digit, a point
/// or a minus), and that is what an engine is asked for here.
#[derive(Clone, Copy)]
pub enum ClassTest<'a> {
    /// **The engine's own test**: given a text and a bracket class — `[0-9]`
    /// or `[^0-9.-]`, the spelling `GLOB`, SQL Server's `LIKE` and every
    /// regular expression share — a condition true where the text holds at
    /// least one character of the class: `x GLOB '*[0-9]*'`,
    /// `x LIKE '%[0-9]%'`, `REGEXP_LIKE(x, '[0-9]')`.
    Flat(&'a dyn Fn(&str, &str) -> String),
    /// **No such test**: the value's digits are folded with `REPLACE` into
    /// one shape and that is asked with `LIKE` alone, about twenty calls deep
    /// over the value. Ask for it only where the engine has neither a bracket
    /// class nor a regular expression.
    Folded,
}

/// **The picture's shape as flat tests** — the rules [`shape_conditions`]
/// states, each asked on its own: the symbol in `LIKE`'s words over the
/// value, and the digits, the point and the minus over the value with its
/// marks out, two of them through the engine's class test.
fn flat_shape_conditions(
    format: &NumberFormat,
    spelling: SidSpelling,
    expr: &str,
    holds: &dyn Fn(&str, &str) -> String,
) -> Vec<String> {
    let prefix = spelling.lower.pattern_literal_prefix;
    let value = normalize_trim(spelling, expr);
    let stripped = strip_marks(format, spelling, expr);
    let not_like = |subject: &str, pattern: &str| {
        format!("{subject} NOT LIKE {prefix}'{}'", pattern.replace('\'', "''"))
    };
    let mut conditions = symbol_conditions(format, spelling, &value);
    conditions.push(holds(&stripped, "[0-9]"));
    conditions.push(format!("NOT {}", holds(&stripped, "[^0-9.-]")));
    let bracketed = format.negatives() == Negatives::Parentheses;
    conditions.push(not_like(&stripped, if bracketed { "%-%" } else { "_%-%" }));
    if format.decimals() > 0 {
        for shape in ["%.%.%", ".%", "-.%", "%."] {
            conditions.push(not_like(&stripped, shape));
        }
    } else {
        conditions.push(not_like(&stripped, "%.%"));
    }
    if !bracketed && !format.opens_with().is_empty() {
        conditions.push(not_like(&value, "_%-%"));
    }
    conditions
}

/// **A symbol at most once, and only at its own end of the value** — after a
/// leading minus, or inside an opening bracket, where the picture writes a
/// negative that way. A symbol holding `%`, `_` or `[` is not held to its
/// place: those are `LIKE`'s own wildcards somewhere.
fn symbol_conditions(format: &NumberFormat, spelling: SidSpelling, value: &str) -> Vec<String> {
    let prefix = spelling.lower.pattern_literal_prefix;
    let like = |pattern: &str| format!("{value} LIKE {prefix}'{}'", pattern.replace('\'', "''"));
    let not_like =
        |pattern: &str| format!("{value} NOT LIKE {prefix}'{}'", pattern.replace('\'', "''"));
    let mut conditions = Vec::new();
    let bracketed = format.negatives() == Negatives::Parentheses;
    let plain = |symbol: &str| !symbol.contains(['%', '_', '['].as_slice());
    let opens = format.opens_with();
    if !opens.is_empty() && plain(opens) {
        let before = if bracketed { "(" } else { "-" };
        conditions.push(not_like(&format!("%{opens}%{opens}%")));
        conditions.push(format!(
            "({} OR {} OR {})",
            not_like(&format!("%{opens}%")),
            like(&format!("{opens}%")),
            like(&format!("{before}{opens}%")),
        ));
    }
    let closes = format.closes_with();
    if !closes.is_empty() && plain(closes) {
        let after = if bracketed { ")" } else { "" };
        conditions.push(not_like(&format!("%{closes}%{closes}%")));
        conditions.push(format!(
            "({} OR {} OR {})",
            not_like(&format!("%{closes}%")),
            like(&format!("%{closes}")),
            like(&format!("%{closes}{after}")),
        ));
    }
    conditions
}

/// **A digit before the first thousands mark** (`fr-1021-02`): `,123` opens
/// with the mark, after nothing or after a leading minus, bracket or symbol,
/// and is not written the picture's way — the pattern-guard engines' shape
/// asks for a digit there, and so does every reading. Nothing where the
/// picture names no thousands mark.
fn thousands_after_a_digit(
    format: &NumberFormat,
    spelling: SidSpelling,
    expr: &str,
) -> Vec<String> {
    if !format.grouped() {
        return Vec::new();
    }
    let prefix = spelling.lower.pattern_literal_prefix;
    let value = normalize_trim(spelling, expr);
    let sign = if format.negatives() == Negatives::Parentheses { "(" } else { "-" };
    let opens = format.opens_with();
    let mut openings = vec![String::new(), sign.to_owned()];
    if !opens.is_empty() && !opens.contains(['%', '_', '['].as_slice()) {
        openings.push(opens.to_owned());
        openings.push(format!("{sign}{opens}"));
    }
    openings
        .iter()
        .map(|opening| {
            format!(
                "{value} NOT LIKE {prefix}'{}'",
                format!("{opening}{THOUSANDS_MARK}%").replace('\'', "''")
            )
        })
        .collect()
}

/// **The picture's shape, as conditions an engine with no pattern match can
/// hold a value to** — `LIKE` and nothing else.
///
/// [`NumberFormat::shape_pattern`] is the shape a pattern-matching engine
/// holds a value to before its marks come out. An engine that converts with a
/// cast that answers nothing, or checks the digits left, used to hold none:
/// the marks came out wherever they stood, so `12$`, `$$12`, `+5` and `.5`
/// read as numbers there while a pattern-matching engine set them aside
/// (`fr-1019-86`). These are the same rules in `LIKE`'s words:
///
/// - a symbol at most once, and only at its own end of the value — after a
///   leading minus, or inside an opening bracket, where the picture writes a
///   negative that way;
/// - with the marks out, nothing left but digits, at most one decimal point
///   and at most a minus, so a plus sign, an exponent or a word is not a
///   number here;
/// - a digit before the decimal point, and at least one after it;
/// - no decimal point at all where the picture has no places;
/// - a minus only as the value's first character, and none where the picture
///   writes a negative in brackets.
///
/// **The value is named as few times as the rules allow**, because it is often
/// a long expression of its own and every mention repeats it. The digits,
/// the point and the minus are asked in one condition ([`canonical_shape`]):
/// a value of the picture's shape comes out as the text `x0` and nothing
/// else. Only a symbol opening the value needs the minus asked again, of the
/// value as written, since `$-12` and `-$12` are the same once it is out.
///
/// A symbol holding `%`, `_` or `[` is not held to its place: those are
/// `LIKE`'s own wildcards somewhere.
fn shape_conditions(format: &NumberFormat, spelling: SidSpelling, expr: &str) -> Vec<String> {
    let prefix = spelling.lower.pattern_literal_prefix;
    let value = normalize_trim(spelling, expr);
    let stripped = strip_marks(format, spelling, expr);
    let mut conditions = symbol_conditions(format, spelling, &value);
    conditions.push(format!("{} LIKE {prefix}'x0'", canonical_shape(format, spelling, &stripped)));
    if format.negatives() != Negatives::Parentheses && !format.opens_with().is_empty() {
        conditions.push(format!("{value} NOT LIKE {prefix}'_%-%'"));
    }
    conditions
}

/// **A value with its marks out, written as the shape it has**: `x0` for one
/// of digits — with a point between two runs of them where the picture has
/// places, and a leading minus where it writes a negative that way — and
/// anything else for any other value.
///
/// Every digit becomes `0`, a run of zeros becomes one, `0.0` becomes `0`, the
/// text is put behind a letter, and a minus straight after the letter comes
/// out: `-1,234.50` is `x0`, while `.5` is left `x.0`, `5.` is `x0.`, `1.2.3`
/// is `x0.0`, `1e3` is `x0e0` and `--5` is `x-0`. Eight halvings take a run of
/// up to 256 digits to one, far past any number a column holds. `REPLACE`,
/// joining and `LIKE` alone, so it reads the same on every engine, a large
/// object included; the letter keeps an engine that reads an empty text as
/// nothing from comparing nothing.
fn canonical_shape(format: &NumberFormat, spelling: SidSpelling, stripped: &str) -> String {
    let replace = spelling.lower.replace;
    let prefix = spelling.lower.pattern_literal_prefix;
    let swap = |text: String, from: &str, to: &str| {
        format!("{replace}({text}, {prefix}'{from}', {prefix}'{to}')")
    };
    let zeros =
        (1..10u8).fold(stripped.to_owned(), |text, digit| swap(text, &digit.to_string(), "0"));
    let run = (0..8).fold(zeros, |text, _| swap(text, "00", "0"));
    let pointed =
        if format.decimals() > 0 { swap(run, &format!("0{DECIMAL_POINT}0"), "0") } else { run };
    let lettered = join_text(spelling, &format!("{prefix}'x'"), &pointed);
    if format.negatives() == Negatives::Parentheses {
        // A bracketed picture writes no minus, so one left is not its shape.
        lettered
    } else {
        swap(lettered, "x-", "x")
    }
}

/// **Two texts joined**, in this engine's own words for joining strings
/// ([`SidSpelling::concat`]).
fn join_text(spelling: SidSpelling, left: &str, right: &str) -> String {
    match spelling.concat {
        Concat::Function(name) | Concat::FunctionBinary(name) => format!("{name}({left}, {right})"),
        Concat::Plus => format!("{left} + {right}"),
        Concat::Operator(operator) => format!("{left} {operator} {right}"),
    }
}

/// **The sign a bracketed negative carries**, as a factor to multiply by.
///
/// A picture that says nothing about brackets needs no factor: the minus sign
/// the source writes survives the marks coming out, and the conversion reads
/// it.
fn bracket_sign(format: &NumberFormat, spelling: SidSpelling, expr: &str) -> Option<String> {
    if format.negatives() != Negatives::Parentheses {
        return None;
    }
    Some(format!("CASE WHEN {} LIKE '(%' THEN -1 ELSE 1 END", normalize_trim(spelling, expr)))
}

/// Multiply the reading by the sign, where the picture asks for one.
fn signed(reading: String, sign: Option<String>) -> String {
    match sign {
        None => reading,
        Some(sign) => format!("({reading}) * ({sign})"),
    }
}

/// **For an engine whose own conversion answers nothing for a value it cannot
/// read** — `TRY_CAST`, `SAFE_CAST`, `TRY_CONVERT` and their neighbours.
///
/// The value is held to the picture's shape first, in `LIKE`'s words and the
/// engine's class test (the rules [`NumberFormat::shape_pattern`] states);
/// then the marks come out and the engine converts what is left. A value
/// outside the shape, or one the conversion cannot read, answers nothing,
/// which is what razeri reads as *this row did not parse*.
///
/// Parameters:
/// - `format`: the picture the design stated;
/// - `spelling`: this engine's vocabulary;
/// - `expr`: an already-rendered text expression;
/// - `safe_cast`: this engine's conversion, given the stripped text;
/// - `class`: how this engine tests a text for a character of a class —
///   [`ClassTest::Flat`] unless it has no such test.
#[must_use]
pub fn read_with_safe_cast(
    format: &NumberFormat,
    spelling: SidSpelling,
    expr: &str,
    safe_cast: &dyn Fn(&str) -> String,
    class: ClassTest<'_>,
) -> String {
    let reading = safe_cast(&strip_marks(format, spelling, expr));
    guarded_reading(format, spelling, expr, &reading, class)
}

/// The reading behind the picture's shape, the thousands mark held in front
/// of the decimal point and behind a digit, and the bracket's sign.
fn guarded_reading(
    format: &NumberFormat,
    spelling: SidSpelling,
    expr: &str,
    reading: &str,
    class: ClassTest<'_>,
) -> String {
    let mut conditions = match class {
        ClassTest::Flat(holds) => flat_shape_conditions(format, spelling, expr, holds),
        ClassTest::Folded => shape_conditions(format, spelling, expr),
    };
    conditions.extend(separators_in_order(format, spelling, expr));
    conditions.extend(thousands_after_a_digit(format, spelling, expr));
    let guarded = format!("CASE WHEN {} THEN {reading} END", conditions.join(" AND "));
    signed(guarded, bracket_sign(format, spelling, expr))
}

/// **For an engine whose own conversion raises, or answers a number for text
/// that is not one** — the shape is checked first, and a value that fails the
/// check is never handed to the conversion at all.
///
/// Parameters:
/// - `format`: the picture the design stated;
/// - `spelling`: this engine's vocabulary;
/// - `expr`: an already-rendered text expression;
/// - `cast`: this engine's conversion, given the stripped text;
/// - `matches`: this engine's own test of a value against a regular
///   expression, given the value and the pattern as a quoted literal.
#[must_use]
pub fn read_with_pattern_guard(
    format: &NumberFormat,
    spelling: SidSpelling,
    expr: &str,
    cast: &dyn Fn(&str) -> String,
    matches: &dyn Fn(&str, &str) -> String,
) -> String {
    let trimmed = normalize_trim(spelling, expr);
    let pattern = format!("'{}'", format.shape_pattern().replace('\'', "''"));
    let reading = cast(&strip_marks(format, spelling, expr));
    // The pattern already holds the thousands mark in front of the decimal
    // point, so this arm needs no second condition for it.
    let guarded = format!("CASE WHEN {} THEN {reading} END", matches(&trimmed, &pattern));
    signed(guarded, bracket_sign(format, spelling, expr))
}

/// **For an engine with neither a conversion that answers nothing nor a way to
/// match a whole pattern** — the value's shape, asked in `LIKE`'s words and
/// the engine's class test, is what says whether it was a number.
///
/// With the picture's marks already out, a value that was a number holds
/// digits, at most one decimal point with a digit either side, and at most a
/// leading minus — the same conditions [`read_with_safe_cast`] holds a value
/// to, and a value that fails them is never handed to the conversion.
///
/// Parameters:
/// - `format`: the picture the design stated;
/// - `spelling`: this engine's vocabulary;
/// - `expr`: an already-rendered text expression;
/// - `cast`: this engine's conversion, given the stripped text;
/// - `class`: how this engine tests a text for a character of a class —
///   [`ClassTest::Flat`] unless it has no such test.
#[must_use]
pub fn read_with_digit_check(
    format: &NumberFormat,
    spelling: SidSpelling,
    expr: &str,
    cast: &dyn Fn(&str) -> String,
    class: ClassTest<'_>,
) -> String {
    let reading = cast(&strip_marks(format, spelling, expr));
    guarded_reading(format, spelling, expr, &reading, class)
}

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

    /// SQL Server's class test, the shape every test here asks with.
    fn like_class(text: &str, class: &str) -> String {
        format!("{text} LIKE '%{class}%'")
    }
    const LIKE_CLASS: fn(&str, &str) -> String = like_class;

    #[test]
    fn a_price_picture_says_a_symbol_thousands_and_two_places() {
        let format = NumberFormat::parse("$#,##0.00").expect("a lawful picture");
        assert_eq!(format.opens_with(), "$");
        assert_eq!(format.closes_with(), "");
        assert!(format.grouped());
        assert_eq!(format.decimals(), 2);
        assert_eq!(format.negatives(), Negatives::Minus);
        assert_eq!(format.marks(), vec!["$".to_owned(), ",".to_owned()]);
    }

    #[test]
    fn a_symbol_may_close_the_value_instead_of_opening_it() {
        let format = NumberFormat::parse("#,##0.00 kr").expect("a lawful picture");
        assert_eq!(format.opens_with(), "");
        assert_eq!(format.closes_with(), " kr");
        assert_eq!(format.marks(), vec![" kr".to_owned(), ",".to_owned()]);
    }

    #[test]
    fn a_plain_picture_names_no_marks_at_all() {
        let format = NumberFormat::parse("0.000").expect("a lawful picture");
        assert!(!format.grouped());
        assert_eq!(format.decimals(), 3);
        assert!(format.marks().is_empty());
        assert_eq!(format.shape_pattern(), "^-?[0-9]+([.][0-9]+)?$");
    }

    #[test]
    fn a_ledger_negative_is_the_same_picture_in_brackets() {
        let format = NumberFormat::parse("$#,##0.00;($#,##0.00)").expect("a lawful picture");
        assert_eq!(format.negatives(), Negatives::Parentheses);
        assert_eq!(
            format.marks(),
            vec!["$".to_owned(), ",".to_owned(), "(".to_owned(), ")".to_owned()]
        );
        assert_eq!(format.spelled(), "$#,##0.00;($#,##0.00)");
    }

    #[test]
    fn the_shape_holds_the_symbol_and_the_groups_where_the_picture_put_them() {
        let format = NumberFormat::parse("$#,##0.00").expect("a lawful picture");
        assert_eq!(format.shape_pattern(), "^-?([$])?[0-9][0-9,]*([.][0-9]+)?$");
    }

    #[test]
    fn the_reading_holds_the_thousands_mark_in_front_of_the_decimal_point() {
        let format = NumberFormat::parse("$#,##0.00").expect("a lawful picture");
        let read = read_with_safe_cast(
            &format,
            crate::ANSI_SID_SPELLING,
            "v",
            &|text| format!("TRY_CAST({text} AS DECIMAL(12, 2))"),
            ClassTest::Folded,
        );
        // Taking the thousands mark out of `2.899,99` leaves `2.89999`, which
        // converts to a price nobody wrote. The condition is what stops it.
        assert!(read.contains("TRIM(v) NOT LIKE '%.%,%'"), "{read}");
    }

    #[test]
    fn a_picture_with_no_thousands_mark_needs_no_such_condition() {
        let format = NumberFormat::parse("0.00").expect("a lawful picture");
        let read = read_with_safe_cast(
            &format,
            crate::ANSI_SID_SPELLING,
            "v",
            &|text| format!("TRY_CAST({text} AS DECIMAL(12, 2))"),
            ClassTest::Folded,
        );
        assert!(!read.contains("'%.%,%'"), "no thousands mark to put in its place: {read}");
        assert!(read.ends_with("THEN TRY_CAST(TRIM(v) AS DECIMAL(12, 2)) END"), "{read}");
    }

    /// **A safe-cast reading holds the value to the picture's shape first**
    /// (`fr-1019-86`): a symbol at most once and at its own end, nothing but
    /// digits, a point and a minus left, a digit either side of the point, a
    /// minus only first. Executed on SQL
    /// Server, `DuckDB` and `SQLite` over `12$`, `$$12`, `+5`, `.5` and `$-12`,
    /// each set aside on all three.
    #[test]
    fn a_safe_cast_reading_holds_the_value_to_the_pictures_shape() {
        let format = NumberFormat::parse("$#,##0.00").expect("a lawful picture");
        let read = read_with_safe_cast(
            &format,
            crate::ANSI_SID_SPELLING,
            "v",
            &|text| format!("TRY_CAST({text} AS DECIMAL(12, 2))"),
            ClassTest::Folded,
        );
        for guard in [
            "TRIM(v) NOT LIKE '%$%$%'",
            "TRIM(v) LIKE '-$%'",
            "'00', '0'), '0.0', '0'))",
            "'x-', 'x') LIKE 'x0'",
            "TRIM(v) NOT LIKE '_%-%'",
        ] {
            assert!(read.contains(guard), "no {guard}: {read}");
        }
        let closing = NumberFormat::parse("#,##0.00 kr").expect("a lawful picture");
        let read = read_with_safe_cast(
            &closing,
            crate::ANSI_SID_SPELLING,
            "v",
            &|text| format!("TRY_CAST({text} AS DECIMAL(12, 2))"),
            ClassTest::Folded,
        );
        assert!(read.contains("TRIM(v) LIKE '% kr'"), "{read}");
    }

    #[test]
    fn a_whole_numbers_digit_check_leaves_no_room_for_a_decimal_point() {
        let whole = NumberFormat::parse("#0").expect("a lawful picture");
        let read = read_with_digit_check(
            &whole,
            crate::ANSI_SID_SPELLING,
            "v",
            &|text| format!("CAST({text} AS INTEGER)"),
            ClassTest::Folded,
        );
        // `1.5` is not a whole number, and an engine with no safe conversion
        // would otherwise answer `1`.
        assert!(read.contains("'x-', 'x') LIKE 'x0'") && !read.contains("'0.0', '0')"), "{read}");
        let fractional = NumberFormat::parse("#0.0").expect("a lawful picture");
        let read = read_with_digit_check(
            &fractional,
            crate::ANSI_SID_SPELLING,
            "v",
            &|text| format!("CAST({text} AS DECIMAL(18, 6))"),
            ClassTest::Folded,
        );
        assert!(read.contains("'0.0', '0')), 'x-', 'x') LIKE 'x0'"), "{read}");
    }

    /// **The value is named a fixed few times, however its shape is asked.**
    /// It is often a long expression of its own — a column read through a
    /// guard of its width — and every mention repeats it, so a check that
    /// named it once per digit doubled a whole load's text.
    #[test]
    fn a_reading_names_the_value_a_fixed_few_times() {
        let value = "COLUMN_VALUE";
        for picture in ["#0", "#0.0", "0.00"] {
            let format = NumberFormat::parse(picture).expect("a lawful picture");
            let read = read_with_safe_cast(
                &format,
                crate::ANSI_SID_SPELLING,
                value,
                &|text| format!("TRY_CAST({text} AS DECIMAL(12, 2))"),
                ClassTest::Folded,
            );
            // The shape and the reading itself.
            assert_eq!(read.matches(value).count(), 2, "{picture}: {read}");
        }
    }

    /// How deep the calls in `sql` nest, string literals left out: the
    /// measure an engine's parser stack holds a reading to.
    fn depth(sql: &str) -> usize {
        let (mut depth, mut deepest, mut quoted) = (0usize, 0usize, false);
        for c in sql.chars() {
            match c {
                '\'' => quoted = !quoted,
                '(' if !quoted => {
                    depth += 1;
                    deepest = deepest.max(depth);
                }
                ')' if !quoted => depth = depth.saturating_sub(1),
                _ => {}
            }
        }
        deepest
    }

    /// **The shared reading is flat** (`fr-1021-01`): a value is no deeper
    /// than its picture's marks' own `REPLACE`s and the test around them, so
    /// an engine with a shallow parser stack reads a wide table. Folded, a
    /// `$#,##0.00` reading nested 23 deep over the value.
    #[test]
    fn the_shared_reading_is_flat() {
        let format = NumberFormat::parse("$#,##0.00").expect("a lawful picture");
        let read = read_with_safe_cast(
            &format,
            crate::ANSI_SID_SPELLING,
            "v",
            &|text| format!("TRY_CAST({text} AS DECIMAL(12, 2))"),
            ClassTest::Flat(&LIKE_CLASS),
        );
        assert!(depth(&read) <= 6, "{} deep: {read}", depth(&read));
    }

    /// **A value that opens with the thousands mark is not a number**
    /// (`fr-1021-02`): `,123` is set aside, as the pattern-guard engines set it
    /// aside, never read as `123`.
    #[test]
    fn a_value_opening_with_the_thousands_mark_is_set_aside() {
        let format = NumberFormat::parse("$#,##0.00").expect("a lawful picture");
        for class in [ClassTest::Flat(&LIKE_CLASS), ClassTest::Folded] {
            let read = read_with_safe_cast(
                &format,
                crate::ANSI_SID_SPELLING,
                "v",
                &|text| format!("TRY_CAST({text} AS DECIMAL(12, 2))"),
                class,
            );
            for opening in ["',%'", "'-,%'", "'$,%'", "'-$,%'"] {
                assert!(read.contains(&format!("TRIM(v) NOT LIKE {opening}")), "{opening}: {read}");
            }
        }
    }

    #[test]
    fn a_bracketed_negative_is_read_as_a_sign_rather_than_a_mark() {
        let format = NumberFormat::parse("$#,##0.00;($#,##0.00)").expect("a lawful picture");
        let read = read_with_safe_cast(
            &format,
            crate::ANSI_SID_SPELLING,
            "v",
            &|text| format!("TRY_CAST({text} AS DECIMAL(12, 2))"),
            ClassTest::Folded,
        );
        assert!(read.contains("CASE WHEN TRIM(v) LIKE '(%' THEN -1 ELSE 1 END"), "{read}");
        assert!(read.contains("REPLACE(REPLACE(REPLACE(REPLACE(TRIM(v)"), "{read}");
    }

    #[test]
    fn a_picture_with_no_digits_is_refused() {
        assert_eq!(NumberFormat::parse("$"), Err(NumberFormatError::NoDigits));
        assert_eq!(NumberFormat::parse(""), Err(NumberFormatError::Empty));
    }

    #[test]
    fn a_source_writing_the_separators_the_other_way_round_is_refused_by_name() {
        let refused = NumberFormat::parse("#.##0,00").expect_err("not covered");
        assert_eq!(refused, NumberFormatError::MarkAfterPoint);
        assert!(refused.to_string().contains("2.899,99"), "{refused}");
    }

    #[test]
    fn a_thousands_mark_separates_groups_of_three() {
        assert_eq!(NumberFormat::parse("#,##0,0"), Err(NumberFormatError::GroupIsNotThree(1)));
    }

    #[test]
    fn two_decimal_points_are_refused() {
        assert_eq!(NumberFormat::parse("0.0.0"), Err(NumberFormatError::TwoPoints));
    }

    #[test]
    fn digits_on_both_sides_of_the_symbol_are_refused() {
        assert_eq!(NumberFormat::parse("0 kr 0"), Err(NumberFormatError::DigitsAfterSymbol));
    }

    #[test]
    fn a_second_picture_that_is_not_the_first_in_brackets_is_refused() {
        assert_eq!(
            NumberFormat::parse("$#,##0.00;-$#,##0.00"),
            Err(NumberFormatError::NegativeIsNotBracketed)
        );
        assert_eq!(NumberFormat::parse("0;(0);0"), Err(NumberFormatError::TooManyPictures));
    }
}