num2words2-core 0.1.3

//! Port of `lang_SR_LATN.py` (Serbian, Gaj's Latin alphabet).
//!
//! Registry check: `CONVERTER_CLASSES["sr_Latn"]` is
//! `lang_SR_LATN.Num2Word_SR_LATN()` (`__init__.py:316`), so the key really
//! does resolve to the class named in the brief.
//!
//! Shape: **self-contained**. `Num2Word_SR_LATN` subclasses `Num2Word_SR`,
//! which subclasses `Num2Word_Base` but defines no `high_numwords` /
//! `mid_numwords` / `low_numwords`. Python therefore never builds
//! `self.cards` and never sets `MAXVAL`, so `cards`/`maxval`/`merge` stay at
//! their trait defaults here and there is **no overflow check**. The only
//! ceiling is the `SCALE` table (10**33 and up raises `KeyError` — see
//! [`scale`]).
//!
//! `Num2Word_SR_LATN` itself is a pure transliteration wrapper: every method
//! calls `super()` and pipes the Cyrillic result through `cyrl_to_latn`. All
//! the arithmetic lives in `lang_SR.py`. This port keeps that structure —
//! tables stay in Cyrillic and transliteration happens at the boundary —
//! because the layering is load-bearing for `to_ordinal` (see below).
//!
//! # Inheritance chain, resolved
//!
//! | method          | defined in       | behaviour                                    |
//! |-----------------|------------------|----------------------------------------------|
//! | `to_cardinal`   | SR_LATN → SR     | `cyrl_to_latn(_int2word(int(n), False))`     |
//! | `to_ordinal`    | SR_LATN → SR     | dict lookup, else cardinal + "и"             |
//! | `to_ordinal_num`| SR_LATN → *Base* | `cyrl_to_latn(value)` → **AttributeError**   |
//! | `to_year`       | SR_LATN → *Base* | `cyrl_to_latn(self.to_cardinal(value))`      |
//!
//! # Faithfully reproduced Python bugs
//!
//! This is a port, not a rewrite. All of the following are verified against
//! the frozen corpus (`bench/corpus.jsonl`, `"lang": "sr_Latn"`):
//!
//! 1. **`to_ordinal_num` raises `AttributeError` for every input.** `SR` and
//!    `SR_LATN` both fail to override it, so it lands on
//!    `Num2Word_Base.to_ordinal_num`, which returns `value` — an `int`, not a
//!    `str`. `SR_LATN.to_ordinal_num` then hands that int to `cyrl_to_latn`,
//!    which immediately calls `s.replace(...)` on it:
//!    `AttributeError: 'int' object has no attribute 'replace'`. **Every**
//!    `ordinal_num` row in the corpus is `{"ok": false, "err":
//!    "AttributeError"}` — there are no successful ones. See
//!    [`attribute_error`] for how the variant is encoded.
//! 2. **`to_ordinal` just glues "и" onto the cardinal** for anything outside
//!    its small lookup dict (Python's own comment: "This is a simplified
//!    implementation"). This produces non-words throughout, and they are the
//!    expected output: `to_ordinal(200)` == "dvestai", `to_ordinal(2000)` ==
//!    "dve hiljadei", `to_ordinal(42)` == "četrdeset dvai",
//!    `to_ordinal(0)` == "nulai", `to_ordinal(10**9)` == "milijardai".
//! 3. **Negative ordinals are cheerfully produced**, unlike most languages
//!    which raise: `to_ordinal(-1)` == "minus jedani". `verify_ordinal` is
//!    never called anywhere in this chain.
//! 4. **`_int2word` drops `feminine` when recursing for negatives**
//!    (`self._int2word(abs(number))` omits the argument, so it silently
//!    resets to `False`). Out of scope for the four modes here — `to_cardinal`
//!    only ever passes `feminine=False` — but reproduced anyway in
//!    [`int2word`] so the bug is preserved if the parameter is ever threaded
//!    through.
//! 5. **`to_ordinal` transliterates twice.** `Num2Word_SR.to_ordinal` calls
//!    `self.to_cardinal(num)`, which dynamically dispatches to the *SR_LATN*
//!    override — so the cardinal is already Latin. It appends the Cyrillic
//!    "и", and the `SR_LATN.to_ordinal` wrapper then re-runs `cyrl_to_latn`
//!    over the whole string. The second pass is a no-op on the Latin part
//!    (no replacement value contains a Cyrillic codepoint, so nothing
//!    cascades) and converts the trailing "и" → "i". Mirrored exactly in
//!    [`LangSrLatn::to_ordinal`] rather than short-circuited.
//! 6. **`to_currency` ignores the currency code entirely for `int` input.**
//!    `Num2Word_SR.to_currency` intercepts `isinstance(val, int)` before
//!    delegating to Base and hardcodes "динар"/"динара", so *every* code —
//!    including ones with no `CURRENCY_FORMS` entry — renders as dinars and
//!    nothing raises. `cents`, `separator` and `adjective` are dropped on that
//!    path too. The corpus pins all of it: `currency:JPY 100` → "sto dinara",
//!    `currency:CHF 1` → "jedan dinar". Only the *float* path reaches
//!    `Num2Word_Base.to_currency` and can raise `NotImplementedError`, which is
//!    why every code has successful int rows and NotImplementedError float rows.
//! 7. **`to_cheque` prints a stringified `bool` where the currency name
//!    belongs.** See [`build_currency_forms`] — this is the reason the forms
//!    table carries a fourth `"True"`/`"False"` element.
//!
//! # Not a bug, but load-bearing: the four-element currency tuples
//!
//! `Num2Word_SR.CURRENCY_FORMS` is the only table in the library whose form
//! tuples end in a `bool` rather than a string. It changes the behaviour of two
//! *inherited* methods it was never designed for; [`build_currency_forms`]
//! documents how the arity is preserved here.
//!
//! # A note on `cyrl_to_latn`
//!
//! The Python comment claims two-character mappings "must come BEFORE
//! single-character entries so they win during the longest-match scan". This
//! is misleading: every *key* is a single codepoint (Љ is U+0409, not Л+Ј),
//! so there is no longest-match scan and the order is irrelevant. The
//! "digraphs" are digraphs on the *output* side ("Lj"). A plain per-char map
//! is therefore exactly equivalent — see [`cyrl_to_latn`].

use crate::base::{Kwargs, KwVal, Lang, N2WError, Result};
use crate::currency::{default_to_currency, CurrencyForms, CurrencyValue};
use crate::floatpath::FloatValue;
use crate::strnum::{python_decimal_parse, python_decimal_str, ParsedNumber};
use num_bigint::BigInt;
use num_integer::Integer;
use num_traits::{FromPrimitive, One, Signed, ToPrimitive, Zero};
use std::collections::HashMap;

const ZERO: &str = "нула";
/// `setup()` sets `self.negword = "минус"` and `self.pointword = "запета"`
/// (the latter is transliterated to "zapeta" and used by the float path —
/// see [`LangSrLatn::to_cardinal_float`]).
const NEGWORD: &str = "минус";

/// `ONES`, keys 1..=9, `(masculine, feminine)`. Index 0 is absent in Python
/// (guarded by `digit_right > 0`).
const ONES: [(&str, &str); 10] = [
    ("", ""), // absent in Python
    ("један", "једна"),
    ("два", "две"),
    ("три", "три"),
    ("четири", "четири"),
    ("пет", "пет"),
    ("шест", "шест"),
    ("седам", "седам"),
    ("осам", "осам"),
    ("девет", "девет"),
];

/// `TENS`, keys 0..=9 — the teens, reached when `digit_mid == 1`.
const TENS: [&str; 10] = [
    "десет",
    "једанаест",
    "дванаест",
    "тринаест",
    "четрнаест",
    "петнаест",
    "шеснаест",
    "седамнаест",
    "осамнаест",
    "деветнаест",
];

/// `TWENTIES`, keys 2..=9. Indices 0 and 1 are absent in Python (guarded by
/// `digit_mid > 1`).
const TWENTIES: [&str; 10] = [
    "", // absent in Python
    "", // absent in Python
    "двадесет",
    "тридесет",
    "четрдесет",
    "педесет",
    "шездесет",
    "седамдесет",
    "осамдесет",
    "деведесет",
];

/// `HUNDREDS`, keys 1..=9. Index 0 is absent in Python (guarded by
/// `digit_left > 0`).
const HUNDREDS: [&str; 10] = [
    "", // absent in Python
    "сто",
    "двеста",
    "триста",
    "четиристо",
    "петсто",
    "шесто",
    "седамсто",
    "осамсто",
    "деветсто",
];

/// `SCALE[chunk_len]` → `(one, few, many, is_feminine)`.
///
/// Python's `SCALE` is a dict with keys 0..=10, i.e. it tops out at 10**30
/// ("квинтилион"). A chunk index of 11 or more — any value >= 10**33 —
/// raises `KeyError`, which is the *only* ceiling this language has (there is
/// no `MAXVAL`, so no `OverflowError`). Serbian uses the long scale, hence
/// милијарда at 10**9 rather than a short-scale "billion".
fn scale(chunk_len: usize) -> Result<(&'static str, &'static str, &'static str, bool)> {
    Ok(match chunk_len {
        0 => ("", "", "", false),
        1 => ("хиљада", "хиљаде", "хиљада", true), // 10^3
        2 => ("милион", "милиона", "милиона", false), // 10^6
        3 => ("милијарда", "милијарде", "милијарди", true), // 10^9 - long scale
        4 => ("билион", "билиона", "билиона", false), // 10^12
        5 => ("билијарда", "билијарде", "билијарди", true), // 10^15
        6 => ("трилион", "трилиона", "трилиона", false), // 10^18
        7 => ("трилијарда", "трилијарде", "трилијарди", true), // 10^21
        8 => ("квадрилион", "квадрилиона", "квадрилиона", false), // 10^24
        9 => ("квадрилијарда", "квадрилијарде", "квадрилијарди", true), // 10^27
        10 => ("квинтилион", "квинтилиона", "квинтилиона", false), // 10^30
        _ => return Err(key_error(&chunk_len.to_string())),
    })
}

// --- Python exception encoding -------------------------------------------

/// Python raised `AttributeError`, which `base.rs` cannot express: there is no
/// `N2WError::Attribute` variant. Following the convention set by
/// `lang_rm_puter.rs` / `lang_rm_sutsilv.rs`, emit `N2WError::Type` carrying a
/// message that names the real exception type so the integration layer can
/// remap it.
///
/// **The bridge must map this back to `AttributeError`, not `TypeError`.**
fn attribute_error(msg: &str) -> N2WError {
    N2WError::Attribute(msg.to_string())
}

/// Python raised `KeyError` — a missing `SCALE` entry for values >= 10**33.
fn key_error(msg: &str) -> N2WError {
    N2WError::Key(msg.to_string())
}

// --- utils.py ------------------------------------------------------------

/// `utils.splitbyx(n, 3)` over a decimal string, with `format_int=True`.
///
/// `start = length % x` guarantees `length - start` is a multiple of `x`, so
/// every subsequent slice is exactly `x` long and `i + x <= length` always
/// holds; the `min` is belt-and-braces against a panic where Python would
/// simply yield a short slice.
///
/// The input is `str(number)` for a **non-negative** number here (`int2word`
/// strips the sign before calling), so it is pure ASCII digits and byte
/// slicing is safe.
fn splitbyx(n: &str, x: usize) -> Vec<u32> {
    let length = n.len();
    let mut out = Vec::new();
    if length > x {
        let start = length % x;
        if start > 0 {
            out.push(n[..start].parse::<u32>().unwrap_or(0));
        }
        let mut i = start;
        while i < length {
            out.push(n[i..(i + x).min(length)].parse::<u32>().unwrap_or(0));
            i += x;
        }
    } else {
        out.push(n.parse::<u32>().unwrap_or(0));
    }
    out
}

/// `utils.get_digits(n)` → `(digit_right, digit_mid, digit_left)`.
///
/// Python: `[int(x) for x in reversed(list(("%03d" % n)[-3:]))]` — zero-pad to
/// three, keep the last three, reverse. Every chunk from `splitbyx(_, 3)` is
/// in 0..=999, so the `[-3:]` truncation never actually bites.
fn get_digits(n: u32) -> (usize, usize, usize) {
    let s = format!("{:03}", n);
    let chars: Vec<char> = s.chars().collect();
    let last3 = &chars[chars.len().saturating_sub(3)..];
    let d: Vec<usize> = last3
        .iter()
        .rev()
        .map(|c| c.to_digit(10).unwrap_or(0) as usize)
        .collect();
    (d[0], d[1], d[2])
}

/// Parse an all-ASCII-digit slice of a reconstructed numeric string (a
/// `left`/`right` fragment from the float path) into a `BigInt`.
///
/// The inputs are internally generated by `format!`/`BigDecimal::with_scale`,
/// so they are always non-empty ASCII digits and the parse cannot fail; the
/// error arm mirrors the `ValueError` Python's `int()` would raise on a bad
/// token rather than panicking. Leading zeros are accepted, matching `int`.
fn parse_bigint(s: &str) -> Result<BigInt> {
    s.parse::<BigInt>()
        .map_err(|_| N2WError::Value(format!("invalid literal for int() with base 10: '{}'", s)))
}

/// Python's `str(number)` for a float/Decimal input — the string
/// `Num2Word_SR.to_cardinal` splits (`n = str(number).replace(",", ".")`).
///
/// * **float** — `repr(float)`. Fixed notation (`{:.*}` at the shim's
///   repr-derived precision) whenever the repr shows a decimal point;
///   otherwise repr picked exponent form (`abs(v) >= 1e16`: "1e+16"),
///   reconstructed by [`python_float_exp_repr`]. The exponent form has no
///   ".", so `to_cardinal` falls to `int("1e+16")` and raises ValueError —
///   corpus-pinned (`cardinal(1e+16)` == ValueError), while
///   `ordinal(1e+16)` (which `int()`s the *value*) succeeds.
/// * **Decimal** — [`python_decimal_str`], the spec `__str__`: trailing
///   zeros kept ("5.00"), scientific form for positive exponents ("1E+2",
///   "1E+20" — again ValueError from `int()`).
///
/// Known gap (unpinned): a float in `(0, 1e-4)` also reprs in exponent form
/// ("1e-05"), but `FloatValue::has_visible_point` reports `true` for every
/// finite fractional float, so such values take the fixed-notation arm here
/// and render instead of raising. No `sr_Latn` corpus row reaches that range.
fn python_number_str(v: &FloatValue) -> String {
    match v {
        FloatValue::Float { value, precision } => {
            if v.has_visible_point() {
                format!("{:.*}", *precision as usize, value)
            } else {
                python_float_exp_repr(*value)
            }
        }
        FloatValue::Decimal { value, .. } => python_decimal_str(value),
    }
}

/// Python's `repr(float)` for the exponent-form cases: shortest mantissa
/// (Rust's `{:e}` is shortest-round-trip, the same contract), "e", an
/// explicit sign and a >= 2-digit exponent — `1e16` -> "1e+16", `1.5e20` ->
/// "1.5e+20". Non-finite values print as Python does ("inf"/"-inf"/"nan");
/// `int()` on those strings raises the same ValueError as on "1e+16".
fn python_float_exp_repr(v: f64) -> String {
    if v.is_nan() {
        return "nan".to_string();
    }
    if v.is_infinite() {
        return if v.is_sign_negative() { "-inf" } else { "inf" }.to_string();
    }
    let s = format!("{:e}", v);
    match s.split_once('e') {
        Some((mant, exp)) => {
            let e: i64 = exp.parse().unwrap_or(0);
            let sign = if e < 0 { "-" } else { "+" };
            format!("{}e{}{:02}", mant, sign, e.abs())
        }
        None => s,
    }
}

/// Extract SR's `feminine=` kwarg (`to_cardinal(self, number,
/// feminine=False)`, which SR_LATN forwards via `*args, **kwargs`). The flag
/// only ever feeds `feminine or SCALE[chunk_len][-1]` then
/// `int(is_feminine)`, so absent / `False` / explicit `None` are all the
/// masculine column and `True` is the feminine one. Any other key, or a
/// non-bool value (`feminine=2` would `int()` to a tuple index Python
/// crashes on), is `NotImplemented` so the dispatcher falls back to the
/// original Python and its genuine semantics.
fn feminine_kwarg(kw: &Kwargs) -> Result<bool> {
    if !kw.only(&["feminine"]) {
        return Err(N2WError::Fallback("kwargs".into()));
    }
    match kw.get("feminine") {
        Option::None | Some(KwVal::None) => Ok(false),
        Some(KwVal::Bool(b)) => Ok(*b),
        Some(_) => Err(N2WError::Fallback("kwargs".into())),
    }
}

// --- lang_SR.py ----------------------------------------------------------

/// `Num2Word_SR.pluralize(number, forms)` — the Slavic 1 / 2-4 / other rule.
fn pluralize(number: u32, forms: (&'static str, &'static str, &'static str, bool)) -> &'static str {
    let form = if number % 100 < 10 || number % 100 > 20 {
        if number % 10 == 1 {
            0
        } else if number % 10 > 1 && number % 10 < 5 {
            1
        } else {
            2
        }
    } else {
        2
    };
    match form {
        0 => forms.0,
        1 => forms.1,
        _ => forms.2,
    }
}

/// `Num2Word_SR._int2word`. Returns **Cyrillic**; the caller transliterates.
fn int2word(number: &BigInt, feminine: bool) -> Result<String> {
    if number.is_negative() {
        // Python: " ".join([self.negword, self._int2word(abs(number))]).
        // Note the dropped `feminine` argument — reproduced verbatim (bug 4).
        return Ok(format!("{} {}", NEGWORD, int2word(&number.abs(), false)?));
    }
    if number.is_zero() {
        return Ok(ZERO.to_string());
    }

    let mut words: Vec<&'static str> = Vec::new();
    let digits = number.to_string();
    let chunks = splitbyx(&digits, 3);
    let mut chunk_len = chunks.len();

    for chunk in &chunks {
        chunk_len -= 1;
        let (digit_right, digit_mid, digit_left) = get_digits(*chunk);

        if digit_left > 0 {
            words.push(HUNDREDS[digit_left]);
        }

        if digit_mid > 1 {
            words.push(TWENTIES[digit_mid]);
        }

        if digit_mid == 1 {
            words.push(TENS[digit_right]);
        } else if digit_right > 0 {
            // Skip 'један' for thousands (1000, 1001, etc.)
            if !(chunk_len > 0 && *chunk == 1) {
                // SCALE is indexed here *before* the `chunk != 0` guard below,
                // so a >= 10**33 chunk whose skip-condition misses raises
                // KeyError at this line in Python. Order preserved.
                let is_feminine = feminine || scale(chunk_len)?.3;
                words.push(if is_feminine {
                    ONES[digit_right].1
                } else {
                    ONES[digit_right].0
                });
            }
        }

        if chunk_len > 0 && *chunk != 0 {
            words.push(pluralize(*chunk, scale(chunk_len)?));
        }
    }

    Ok(words.join(" "))
}

/// `Num2Word_SR.to_cardinal(number, feminine)` for a float/Decimal `number`
/// — the whole method, both branches, returning **Cyrillic** (the caller
/// transliterates, exactly like the `cyrl_to_latn(super().to_cardinal(...))`
/// wrapper):
///
/// ```python
/// n = str(number).replace(",", ".")
/// if "." in n:
///     is_negative = n.startswith("-")
///     left, right = (n[1:] if is_negative else n).split(".")
///     leading = len(right) - len(right.lstrip("0"))
///     decimal_part = ("нула " * leading) + self._int2word(int(right), feminine)
///     result = "%s %s %s" % (self._int2word(int(left), feminine),
///                            self.pointword, decimal_part)
///     if is_negative: result = self.negword + " " + result
/// else:
///     result = self._int2word(int(n), feminine)
/// return result
/// ```
///
/// The fraction is one whole number, not a digit sequence (`2.675` reads
/// "675"), and every leading fraction zero becomes a "нула" word — `right ==
/// "0"` (from `1.0`) counts as one leading zero *and* `int("0") == 0`, hence
/// the doubled "нула нула"; `Decimal("5.00")`'s `right == "00"` triples it.
/// The `else` arm is where exponent-form strings die: `int("1e+16")` /
/// `int("1E+2")` raise ValueError with Python's exact message (see
/// [`parse_bigint`]).
fn cardinal_float_str(value: &FloatValue, feminine: bool) -> Result<String> {
    // n = str(number).replace(",", ".") — the reconstruction never emits a
    // comma; the sign stays on the string so `startswith("-")` matches.
    let n = python_number_str(value);

    if let Some((left, right)) = {
        let is_negative = n.starts_with('-');
        let abs_n = if is_negative { &n[1..] } else { &n[..] };
        abs_n.split_once('.').map(|(l, r)| (l.to_string(), r.to_string()))
    } {
        let is_negative = n.starts_with('-');
        // leading_zero_count = len(right) - len(right.lstrip("0"))
        let leading = right.chars().take_while(|&c| c == '0').count();
        let left_int = parse_bigint(&left)?;
        let right_int = parse_bigint(&right)?;

        // decimal_part = ("нула " * leading) + _int2word(int(right), feminine)
        let mut decimal_part = String::new();
        for _ in 0..leading {
            decimal_part.push_str(ZERO);
            decimal_part.push(' ');
        }
        decimal_part.push_str(&int2word(&right_int, feminine)?);

        // result = "<int(left)> запета <decimal_part>"
        let body = format!(
            "{} {} {}",
            int2word(&left_int, feminine)?,
            "запета",
            decimal_part
        );
        Ok(if is_negative {
            format!("{} {}", NEGWORD, body)
        } else {
            body
        })
    } else {
        // else: _int2word(int(n), feminine). int(n) keeps the sign, and
        // int2word renders "минус ..." for a negative; exponent-form and
        // non-finite strings ("1e+16", "1E+2", "inf") raise ValueError here.
        int2word(&parse_bigint(&n)?, feminine)
    }
}

/// The ordinal lookup table from `Num2Word_SR.to_ordinal` (Cyrillic).
///
/// Sparse on purpose: 1..=20, the round tens, 100 and 1000. Everything else
/// falls through to the cardinal + "и" path (bug 2).
fn ordinal_word(value: &BigInt) -> Option<&'static str> {
    let n = value.to_u32()?;
    Some(match n {
        1 => "први",
        2 => "други",
        3 => "трећи",
        4 => "четврти",
        5 => "пети",
        6 => "шести",
        7 => "седми",
        8 => "осми",
        9 => "девети",
        10 => "десети",
        11 => "једанаести",
        12 => "дванаести",
        13 => "тринаести",
        14 => "четрнаести",
        15 => "петнаести",
        16 => "шеснаести",
        17 => "седамнаести",
        18 => "осамнаести",
        19 => "деветнаести",
        20 => "двадесети",
        30 => "тридесети",
        40 => "четрдесети",
        50 => "педесети",
        60 => "шездесети",
        70 => "седамдесети",
        80 => "осамдесети",
        90 => "деведесети",
        100 => "стоти",
        1000 => "хиљадити",
        _ => return None,
    })
}

// --- lang_SR_LATN.py -----------------------------------------------------

/// `_CYRL_TO_LATN` / `cyrl_to_latn` — Serbian Cyrillic → Gaj's Latin.
///
/// Python chains 36 `str.replace` calls. Each key is a single codepoint and no
/// replacement *value* contains a Cyrillic codepoint, so no replacement can be
/// re-matched by a later pass — the chain is exactly a per-character map, and
/// running it twice is idempotent (which `to_ordinal` and `to_year` both rely
/// on; see bug 5).
///
/// All 30 letters of the Serbian Cyrillic alphabet are covered. Anything else
/// (spaces, Latin text, digits) passes through untouched, matching `replace`.
fn cyrl_to_latn(s: &str) -> String {
    let mut out = String::with_capacity(s.len());
    for c in s.chars() {
        match c {
            // Digraphs are case-sensitive: one Cyrillic char → two Latin.
            'Љ' => out.push_str("Lj"),
            'Њ' => out.push_str("Nj"),
            'Џ' => out.push_str("Dž"),
            'љ' => out.push_str("lj"),
            'њ' => out.push_str("nj"),
            'џ' => out.push_str("dž"),
            // Single characters.
            'А' => out.push('A'),
            'а' => out.push('a'),
            'Б' => out.push('B'),
            'б' => out.push('b'),
            'В' => out.push('V'),
            'в' => out.push('v'),
            'Г' => out.push('G'),
            'г' => out.push('g'),
            'Д' => out.push('D'),
            'д' => out.push('d'),
            'Ђ' => out.push('Đ'),
            'ђ' => out.push('đ'),
            'Е' => out.push('E'),
            'е' => out.push('e'),
            'Ж' => out.push('Ž'),
            'ж' => out.push('ž'),
            'З' => out.push('Z'),
            'з' => out.push('z'),
            'И' => out.push('I'),
            'и' => out.push('i'),
            'Ј' => out.push('J'),
            'ј' => out.push('j'),
            'К' => out.push('K'),
            'к' => out.push('k'),
            'Л' => out.push('L'),
            'л' => out.push('l'),
            'М' => out.push('M'),
            'м' => out.push('m'),
            'Н' => out.push('N'),
            'н' => out.push('n'),
            'О' => out.push('O'),
            'о' => out.push('o'),
            'П' => out.push('P'),
            'п' => out.push('p'),
            'Р' => out.push('R'),
            'р' => out.push('r'),
            'С' => out.push('S'),
            'с' => out.push('s'),
            'Т' => out.push('T'),
            'т' => out.push('t'),
            'Ћ' => out.push('Ć'),
            'ћ' => out.push('ć'),
            'У' => out.push('U'),
            'у' => out.push('u'),
            'Ф' => out.push('F'),
            'ф' => out.push('f'),
            'Х' => out.push('H'),
            'х' => out.push('h'),
            'Ц' => out.push('C'),
            'ц' => out.push('c'),
            'Ч' => out.push('Č'),
            'ч' => out.push('č'),
            'Ш' => out.push('Š'),
            'ш' => out.push('š'),
            other => out.push(other),
        }
    }
    out
}

// --- currency ------------------------------------------------------------

/// `Num2Word_SR.CURRENCY_FORMS`, transcribed with its **four**-element tuples
/// intact.
///
/// ```python
/// CURRENCY_FORMS = {
///     "RUB": (("rublja", "rublje", "rublji", True),
///             ("kopejka", "kopejke", "kopejki", True)),
///     "EUR": (("evro", "evra", "evra", False),
///             ("cent", "centa", "centi", False)),
///     "RSD": (("dinar", "dinara", "dinara", False),
///             ("para", "pare", "para", True)),
/// }
/// ```
///
/// Two things about this table are unusual and both are load-bearing.
///
/// **The values are already Latin.** `lang_SR.py` keeps its *number* words in
/// Cyrillic but wrote the currency nouns in Latin, so `cyrl_to_latn` passes
/// over them untouched and Cyrillic `sr` inherits Latin currency names. Not our
/// bug to fix; do not "restore" them to Cyrillic.
///
/// **The fourth element is a `bool`, not a string.** It is the grammatical
/// gender flag `_cents_verbose` hands to `_int2word` as its `feminine`
/// argument, via `CURRENCY_FORMS[currency][1][-1]` — para and kopejka are
/// feminine, cent is not. [`CurrencyForms`] can only hold strings, so the flag
/// is carried as the literal `"True"` / `"False"`. That is not a workaround: it
/// is *exactly* the text Python produces wherever that slot reaches a format
/// string. Two consumers depend on the fourth element, and each is satisfied by
/// this encoding:
///
/// 1. `cents_verbose` reads the **subunit's** last element back as a bool
///    (`== "True"`), reproducing `[1][-1]`. Hence
///    `to_currency(1.21, "RSD")` → "jedan dinar, dvadeset **jedna** para"
///    (feminine), while `to_currency(1.21, "EUR")` → "...dvadeset **jedan**
///    cent" (masculine).
/// 2. `Num2Word_Base.to_cheque` does
///    `unit = cr1[-1] if isinstance(cr1, tuple) else cr1`, meaning that for a
///    4-tuple it picks the **bool** instead of the plural noun and interpolates
///    it: `"%s AND %s %s" % (words, "56/100", False)`, then `.upper()`. The
///    corpus pins the result verbatim:
///
///    ```text
///    cheque:EUR 1234.56 -> "HILJADA DVESTA TRIDESET ČETIRI AND 56/100 FALSE"
///    cheque:RUB -2.00   -> "MINUS DVA AND 00/100 TRUE"
///    ```
///
///    RUB yielding "TRUE" where EUR yields "FALSE" is what proves the slot is
///    read positionally, not by type. Storing `"False"`/`"True"` puts the right
///    string where `currency::default_to_cheque`'s `forms.unit.last()` looks,
///    so `to_cheque` needs no override at all.
///
/// **Truncating these to three forms would compile, pass a reading, and
/// silently emit "... AND 56/100 EVRA" — a corpus failure.** `pluralize` only
/// ever indexes 0..=2, so the fourth element never leaks into the plural path.
fn build_currency_forms() -> HashMap<&'static str, CurrencyForms> {
    let mut m: HashMap<&'static str, CurrencyForms> = HashMap::new();
    m.insert(
        "RUB",
        CurrencyForms::new(
            &["rublja", "rublje", "rublji", "True"],
            &["kopejka", "kopejke", "kopejki", "True"],
        ),
    );
    m.insert(
        "EUR",
        CurrencyForms::new(
            &["evro", "evra", "evra", "False"],
            &["cent", "centa", "centi", "False"],
        ),
    );
    m.insert(
        "RSD",
        CurrencyForms::new(
            &["dinar", "dinara", "dinara", "False"],
            &["para", "pare", "para", "True"],
        ),
    );
    m
}

pub struct LangSrLatn {
    /// Built once in [`LangSrLatn::new`]. `lib.rs` holds the language in a
    /// `OnceLock`, so this table is constructed a single time per process and
    /// `to_currency` only ever reads it.
    currency_forms: HashMap<&'static str, CurrencyForms>,
}

impl LangSrLatn {
    pub fn new() -> Self {
        LangSrLatn {
            currency_forms: build_currency_forms(),
        }
    }
}

impl Default for LangSrLatn {
    fn default() -> Self {
        LangSrLatn::new()
    }
}

impl Lang for LangSrLatn {
    fn negword(&self) -> &str {
        NEGWORD
    }

    fn pointword(&self) -> &str {
        "запета"
    }

    /// `SR_LATN.to_cardinal` → `cyrl_to_latn(SR.to_cardinal(...))`.
    ///
    /// `SR.to_cardinal` does `str(number).replace(",", ".")` and branches on a
    /// "." — integer input never contains one, so this reduces to
    /// `_int2word(int(n), feminine=False)`.
    fn to_cardinal(&self, value: &BigInt) -> Result<String> {
        Ok(cyrl_to_latn(&int2word(value, false)?))
    }

    /// `SR_LATN.to_ordinal` → `cyrl_to_latn(SR.to_ordinal(...))`.
    ///
    /// The `try: int(number) except (ValueError, TypeError): return
    /// str(number)` guard in Python can never fire for the BigInt we are
    /// handed, so it is elided.
    fn to_ordinal(&self, value: &BigInt) -> Result<String> {
        if let Some(word) = ordinal_word(value) {
            return Ok(cyrl_to_latn(word));
        }
        // Python: `cardinal = self.to_cardinal(num)` dispatches to the SR_LATN
        // override, so this is already Latin; "и" is Cyrillic; the outer
        // wrapper transliterates the concatenation. Bug 5 — kept verbatim.
        let cardinal = self.to_cardinal(value)?;
        Ok(cyrl_to_latn(&format!("{}{}", cardinal, "и")))
    }

    /// Neither `SR_LATN` nor `SR` overrides `to_ordinal_num`, so it resolves to
    /// `Num2Word_Base.to_ordinal_num`, which returns `value` **unchanged** — an
    /// `int`. `SR_LATN.to_ordinal_num` then feeds that int to `cyrl_to_latn`,
    /// whose first statement is `s.replace("Љ", "Lj")`:
    ///
    /// ```text
    /// AttributeError: 'int' object has no attribute 'replace'
    /// ```
    ///
    /// Unconditional — the corpus has zero successful `ordinal_num` rows for
    /// this language. Bug 1.
    fn to_ordinal_num(&self, _value: &BigInt) -> Result<String> {
        Err(attribute_error(
            "'int' object has no attribute 'replace'",
        ))
    }

    /// `SR_LATN.to_year` → `cyrl_to_latn(Num2Word_Base.to_year(...))`, and
    /// `Base.to_year` is just `self.to_cardinal(value)` — which dispatches back
    /// to the SR_LATN override, so the string is already Latin and the outer
    /// `cyrl_to_latn` is a no-op. No era handling, no two-digit splitting.
    fn to_year(&self, value: &BigInt) -> Result<String> {
        Ok(cyrl_to_latn(&self.to_cardinal(value)?))
    }

    /// `SR_LATN.to_cardinal(number)` for **non-integer** input — the float and
    /// Decimal cardinal path.
    ///
    /// `Num2Word_SR` does **not** override `to_cardinal_float`. Instead its
    /// `to_cardinal` services floats/Decimals inline by string-splitting
    /// `str(number)` (`lang_SR.py`), and the `num2words` dispatcher routes float
    /// input through `converter.to_cardinal(number)` — never through
    /// `to_cardinal_float`. So the inherited `default_to_cardinal_float` is the
    /// *wrong* behaviour for this language (it renders the fractional part
    /// digit-by-digit and leaves the Cyrillic pointword untransliterated, e.g.
    /// `2.675` → "dva запета šest sedam pet"); [`cardinal_float_str`]
    /// reproduces the SR string algorithm instead, and this wrapper is
    /// `cyrl_to_latn(Num2Word_SR.to_cardinal(number))`.
    ///
    /// `precision_override` (issue #580) is a **documented no-op**:
    /// `SR.to_cardinal` never consults `self.precision`, so
    /// `num2words(2.675, lang="sr_Latn", precision=1)` is unchanged — verified
    /// against the live interpreter. This hook also serves the
    /// fractional-cents currency path (`cardinal_from_decimal` ->
    /// `cardinal_from_bigdecimal` dispatches through the trait), mirroring
    /// Python's virtual `self.to_cardinal(float(right))`.
    fn to_cardinal_float(
        &self,
        value: &FloatValue,
        _precision_override: Option<u32>,
    ) -> Result<String> {
        Ok(cyrl_to_latn(&cardinal_float_str(value, false)?))
    }

    /// `to_cardinal(float/Decimal)` — the FULL entry, whole values included.
    ///
    /// SR branches on `"." in str(number)`, so a whole value with a visible
    /// point still takes the float grammar: `5.0` -> "pet zapeta nula nula",
    /// `Decimal("5.00")` -> "pet zapeta nula nula nula", `-0.0` -> "minus
    /// nula zapeta nula nula". A pointless string falls to `int(n)`:
    /// `Decimal("5")` -> "pet", but `1e+16` / `Decimal("1E+2")` raise
    /// ValueError. The base default (whole -> int path) would get every one
    /// of those wrong, hence this override.
    fn cardinal_float_entry(
        &self,
        value: &FloatValue,
        _precision_override: Option<u32>,
    ) -> Result<String> {
        Ok(cyrl_to_latn(&cardinal_float_str(value, false)?))
    }

    /// `to_ordinal(float/Decimal)`. `SR.to_ordinal` opens with
    /// `num = int(number)` — `int()` of the *value*, truncation toward zero —
    /// so `2.5` -> 2 -> "drugi", `-1.5` -> -1 -> "minus jedani", `0.5` ->
    /// 0 -> "nulai", and `1e+16` *succeeds* ("deset bilijardii") where the
    /// cardinal raises ValueError.
    ///
    /// The `except (ValueError, TypeError): return str(number)` guard can
    /// only fire here for `nan` (`int(nan)` is ValueError, so Python returns
    /// `str(nan)` == "nan", which `cyrl_to_latn` passes through); `int(inf)`
    /// raises OverflowError, which the guard does **not** catch. Neither has
    /// a corpus row; both are mirrored for completeness.
    fn ordinal_float_entry(&self, value: &FloatValue) -> Result<String> {
        let num = match value {
            FloatValue::Float { value: f, .. } => {
                if f.is_nan() {
                    return Ok("nan".to_string());
                }
                BigInt::from_f64(f.trunc()).ok_or_else(|| {
                    N2WError::Overflow("cannot convert float infinity to integer".into())
                })?
            }
            // int(Decimal) truncates toward zero; with_scale(0) drops the
            // fractional digits the same way (see floatpath::float2tuple).
            FloatValue::Decimal { value: d, .. } => d.with_scale(0).as_bigint_and_exponent().0,
        };
        // Dispatches to the SR_LATN to_ordinal override above — table word or
        // Latin cardinal + "и", double-transliterated exactly as Python.
        self.to_ordinal(&num)
    }

    /// `to_ordinal_num(float/Decimal)` — bug 1, float edition. Base's
    /// `to_ordinal_num` returns the value **unchanged** (a `float` or a
    /// `Decimal`), and `SR_LATN.to_ordinal_num` hands it to `cyrl_to_latn`,
    /// whose first statement is `s.replace("Љ", "Lj")`:
    ///
    /// ```text
    /// AttributeError: 'float' object has no attribute 'replace'
    /// AttributeError: 'decimal.Decimal' object has no attribute 'replace'
    /// ```
    ///
    /// Unconditional — every `ordinal_num` row in the corpus (int, float and
    /// Decimal alike) is an AttributeError. Only the message differs by type.
    fn ordinal_num_float_entry(&self, value: &FloatValue, _repr_str: &str) -> Result<String> {
        let type_name = match value {
            FloatValue::Float { .. } => "float",
            FloatValue::Decimal { .. } => "decimal.Decimal",
        };
        Err(attribute_error(&format!(
            "'{}' object has no attribute 'replace'",
            type_name
        )))
    }

    // `year_float_entry` is deliberately NOT overridden: `SR_LATN.to_year` is
    // `cyrl_to_latn(Base.to_year(...))` == `cyrl_to_latn(self.to_cardinal(...))`,
    // and the trait default routes through the overridden
    // `cardinal_float_entry` above — so `to_year(5.0)` == "pet zapeta nula
    // nula" and `to_year(1e+16)` raises ValueError, as pinned.

    /// `converter.str_to_number` — Base's `Decimal(value)`, which neither SR
    /// nor SR_LATN overrides. The `Inf` interception reproduces what happens
    /// *next* on the pinned path: `to_cardinal(Decimal("Infinity"))` reads
    /// `n = str(number)` == "Infinity", finds no ".", and dies in
    /// `int("Infinity")` with ValueError. The binding otherwise maps
    /// `ParsedNumber::Inf` to the base integer path's OverflowError before
    /// any SR code runs, so the ValueError must be raised here. (NaN needs no
    /// interception: the binding's ValueError already matches `int("NaN")`'s
    /// type.)
    ///
    /// Known gap (unpinned): Python's `to_ordinal(Decimal("Infinity"))`
    /// raises OverflowError (`int()` of the Decimal *object*), which this
    /// entry-level interception turns into the cardinal path's ValueError.
    /// The strings corpus pins Infinity under `to=cardinal` only.
    fn str_to_number(&self, s: &str) -> Result<ParsedNumber> {
        match python_decimal_parse(s)? {
            ParsedNumber::Inf { negative } => Err(N2WError::Value(format!(
                "invalid literal for int() with base 10: '{}Infinity'",
                if negative { "-" } else { "" }
            ))),
            other => Ok(other),
        }
    }

    // ---- grammatical kwargs ----------------------------------------------

    /// `SR_LATN.to_cardinal(*args, **kwargs)` forwards SR's `feminine=`
    /// kwarg: `to_cardinal(2, feminine=True)` == "dve". The negative
    /// recursion inside `_int2word` still drops the flag (bug 4):
    /// `to_cardinal(-5, feminine=True)` == "minus pet".
    fn to_cardinal_kw(&self, value: &BigInt, kw: &Kwargs) -> Result<String> {
        let feminine = feminine_kwarg(kw)?;
        Ok(cyrl_to_latn(&int2word(value, feminine)?))
    }

    /// The float/Decimal side of the same kwarg: Python's `to_cardinal`
    /// threads `feminine` into *both* `_int2word` calls of the "." branch.
    fn to_cardinal_float_kw(
        &self,
        value: &FloatValue,
        _precision_override: Option<u32>,
        kw: &Kwargs,
    ) -> Result<String> {
        let feminine = feminine_kwarg(kw)?;
        Ok(cyrl_to_latn(&cardinal_float_str(value, feminine)?))
    }

    // ---- currency -------------------------------------------------------
    //
    // `Num2Word_SR` overrides `to_currency`, `pluralize` and `_cents_verbose`;
    // `Num2Word_SR_LATN` wraps `to_currency` in `cyrl_to_latn`. Everything else
    // on the currency surface is inherited from `Num2Word_Base` unchanged and
    // is already mirrored by the trait defaults, so it is deliberately absent
    // here:
    //
    // * `_money_verbose` → `self.to_cardinal(number)`, which dispatches to the
    //   SR_LATN override and so already returns Latin.
    // * `_cents_terse`   → `CURRENCY_PRECISION.get(currency, 100)` is 100 for
    //   every code, so the default's `"%02d"` width is right.
    // * `CURRENCY_PRECISION` is `{}` (never rebound), so `currency_precision`
    //   stays at the trait's flat 100 — SR has no 3-decimal and no 0-decimal
    //   currency, and `default_to_currency`'s `divisor == 1` branch is
    //   unreachable. This is *why* `currency:JPY 12.34` raises
    //   NotImplementedError instead of rounding to a whole yen.
    // * `CURRENCY_ADJECTIVES` is `{}`, so `currency_adjective` stays `None` and
    //   `adjective=True` is a no-op — matching `if adjective and currency in
    //   self.CURRENCY_ADJECTIVES`.
    // * `to_cheque` is Base's; see `build_currency_forms` for why the default
    //   reproduces its bool-instead-of-noun bug without an override.

    /// `self.__class__.__name__` for the NotImplementedError message. The
    /// raise sites live in `base.py` and `lang_SR.py`, but the *instance* is a
    /// `Num2Word_SR_LATN`, so that is the name Python interpolates.
    fn lang_name(&self) -> &str {
        "Num2Word_SR_LATN"
    }

    /// `Num2Word_SR.to_currency(val, currency="RSD", ...)`.
    ///
    /// This override is hand-added rather than generated. The generator reads
    /// `inspect.signature(converter.to_currency)`, and `Num2Word_SR_LATN`'s is
    /// `(*args, **kwargs)` — the real defaults sit one level up on
    /// `Num2Word_SR.to_currency`, invisible to `signature()`. sr_Latn is the
    /// only affected language whose default is not Base's "EUR", so it is the
    /// only one where the omission is observable:
    ///
    /// ```text
    /// num2words(1.5, lang="sr_Latn", to="currency")
    ///   == "jedan dinar, pedeset para"      # RSD, verified against Python
    ///   != "jedan evro, pedeset centi"      # what Base's "EUR" default gives
    /// ```
    ///
    /// (`separator` needs no such override: SR's default is `","`, identical to
    /// Base's, so the trait default is already correct.)
    fn default_currency(&self) -> &str {
        "RSD"
    }

    fn currency_forms(&self, code: &str) -> Option<&CurrencyForms> {
        self.currency_forms.get(code)
    }

    /// `Num2Word_SR.pluralize(number, forms)` — the Slavic one / few / many
    /// rule:
    ///
    /// ```python
    /// if number % 100 < 10 or number % 100 > 20:
    ///     if number % 10 == 1:      form = 0
    ///     elif 1 < number % 10 < 5: form = 1
    ///     else:                     form = 2
    /// else:
    ///     form = 2
    /// return forms[form]
    /// ```
    ///
    /// Distinct from the module-level `pluralize` helper, which serves
    /// `SCALE`'s 4-tuples from inside [`int2word`]. This one is the trait hook
    /// `currency::default_to_currency` calls with `CURRENCY_FORMS` entries.
    ///
    /// `mod_floor` rather than `%` because Python's `%` floors on negatives.
    /// Every caller hands us a non-negative `left`/`right` straight out of
    /// `parse_currency_parts`, so the two agree today — the port simply should
    /// not depend on that.
    fn pluralize(&self, n: &BigInt, forms: &[String]) -> Result<String> {
        let ten = BigInt::from(10);
        let m100 = n.mod_floor(&BigInt::from(100));
        let m10 = n.mod_floor(&ten);

        let form = if m100 < ten || m100 > BigInt::from(20) {
            if m10.is_one() {
                0
            } else if m10 > BigInt::one() && m10 < BigInt::from(5) {
                1
            } else {
                2
            }
        } else {
            2
        };

        // Python indexes the tuple directly. Every SR entry carries four forms
        // so this cannot miss, but a short tuple would raise IndexError rather
        // than panic, and the exception type is what a caller can observe.
        forms
            .get(form)
            .cloned()
            .ok_or_else(|| N2WError::Index("tuple index out of range".into()))
    }

    /// `Num2Word_SR._cents_verbose`:
    /// `self._int2word(number, self.CURRENCY_FORMS[currency][1][-1])`.
    ///
    /// Returns **Cyrillic**, exactly as Python does: `_cents_verbose` is only
    /// ever reached from inside `to_currency`, whose SR_LATN wrapper
    /// transliterates the finished string. Nothing else calls this hook, and
    /// the bindings do not expose it, so the Cyrillic cannot escape.
    ///
    /// `[1][-1]` is the *subunit* tuple's gender flag — see
    /// [`build_currency_forms`] for why it round-trips through `"True"`.
    fn cents_verbose(&self, number: &BigInt, currency: &str) -> Result<String> {
        // Python would raise KeyError here for an unknown code, but the lookup
        // is unreachable: `default_to_currency` resolves the same table first
        // and has already raised NotImplementedError by this point.
        let forms = self
            .currency_forms
            .get(currency)
            .ok_or_else(|| key_error(currency))?;
        let feminine = forms.subunit.last().map(|s| s == "True").unwrap_or(false);
        int2word(number, feminine)
    }

    /// `Num2Word_SR_LATN.to_currency` →
    /// `cyrl_to_latn(Num2Word_SR.to_currency(*args, **kwargs))`.
    ///
    /// `Num2Word_SR.to_currency` splits in two: it services `int` itself and
    /// hands everything else to `Num2Word_Base.to_currency` via `super()`. The
    /// int arm is bug 6 — it never looks at `currency`.
    fn to_currency(
        &self,
        val: &CurrencyValue,
        currency: &str,
        cents: bool,
        separator: Option<&str>,
        adjective: bool,
    ) -> Result<String> {
        // The trait hands us None when the caller omitted `separator=`;
        // resolve it before the ported body. SR's default is Base's ",".
        let separator = separator.unwrap_or(self.default_separator());

        if let CurrencyValue::Int(v) = val {
            // parse_currency_parts(val, is_int_with_cents=False) reduces to
            // `(abs(val), 0, val < 0)` — the divisor is never consulted.
            let is_negative = v.is_negative();
            let left = v.abs();

            let mut words: Vec<String> = Vec::new();
            if is_negative {
                // Python hardcodes this literal rather than using self.negword.
                // They happen to be the same string; kept as the literal so the
                // two stay independent if negword ever moves.
                words.push("минус".to_string());
            }
            // `self.to_cardinal(left)` dispatches to the SR_LATN override, so
            // this fragment arrives already Latin and the outer cyrl_to_latn
            // re-runs over it idempotently — bug 5's shape again. `left` is
            // non-negative, so `_int2word` never prepends its own "минус".
            // A `left >= 10**33` still raises KeyError from `scale`, exactly as
            // Python does: the int arm has no ceiling of its own.
            words.push(self.to_cardinal(&left)?);

            // Python's `elif 2 <= left % 10 <= 4 and not (12 <= left % 100 <= 14)`
            // and its `else` both append "динара", so the elif is dead and the
            // whole cascade collapses to this. Note 11 -> "динара" (the
            // `% 100 != 11` guard) but 21 -> "динар".
            let one_dinar = left.mod_floor(&BigInt::from(10)).is_one()
                && left.mod_floor(&BigInt::from(100)) != BigInt::from(11);
            let unit = if one_dinar {
                "динар"
            } else {
                "динара"
            };
            words.push(unit.to_string());

            // `currency`, `cents`, `separator` and `adjective` are all dropped
            // on this path — an unknown code raises nothing and still says
            // dinars. Bug 6.
            let _ = (currency, cents, separator, adjective);
            return Ok(cyrl_to_latn(&words.join(" ")));
        }

        // Floats/Decimals: `super().to_currency(...)` — Num2Word_Base's, which
        // is what raises NotImplementedError for a code outside the table.
        // The `?` propagates that error untransliterated, as Python's raise
        // unwinds past `cyrl_to_latn` without calling it.
        let out = default_to_currency(self, val, currency, cents, separator, adjective)?;
        Ok(cyrl_to_latn(&out))
    }
}