tgen.h

/*
 * Copyright (c) 2026 Bruno Monteiro
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to deal
 * in the Software without restriction, including without limitation the rights
 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 * copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in
 * all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 * THE SOFTWARE.
 */
 
#pragma once
 
#include <algorithm>
#include <bitset>
#include <cstdint>
#include <functional>
#include <initializer_list>
#include <iomanip>
#include <iostream>
#include <limits>
#include <map>
#include <optional>
#include <queue>
#include <random>
#include <set>
#include <sstream>
#include <stdexcept>
#include <string>
#include <sys/types.h>
#include <type_traits>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
 
namespace tgen {
 
/**************************
 *                        *
 *   GENERAL OPERATIONS   *
 *                        *
 **************************/
 
namespace detail {
 
// Type aliases.
using u128 = unsigned __int128;
using i128 = __int128;
 
/*
 * Error handling.
 */
 
inline void throw_assertion_error(const std::string &condition,
                                  const std::string &msg, const char *file,
                                  int line) {
    throw std::runtime_error("tgen: " + msg + " (assertion `" + condition +
                             "` failed at " + file + ":" +
                             std::to_string(line) + ")");
}
inline void throw_assertion_error(const std::string &condition,
                                  const char *file, int line) {
    throw std::runtime_error("tgen: assertion `" + condition + "` failed at " +
                             std::string(file) + ":" + std::to_string(line));
}
inline std::runtime_error error(const std::string &msg) {
    return std::runtime_error("tgen: " + msg);
}
inline std::runtime_error contradiction_error(const std::string &type,
                                              const std::string &msg = "") {
    // Tried to generate a contradictory type.
    std::string error_msg =
        type + ": invalid " + type + " (contradictory restrictions)";
    if (!msg.empty())
        error_msg += ": " + msg;
    return error(error_msg);
}
inline std::runtime_error
complex_restrictions_error(const std::string &type,
                           const std::string &msg = "") {
    // Tried to generate a type with too many distinct restrictions.
    std::string error_msg =
        type + ": cannot represent " + type + " (complex restrictions)";
    if (!msg.empty())
        error_msg += ": " + msg;
    return error(error_msg);
}
inline void tgen_ensure_against_bug(bool cond, const std::string &msg = "") {
    if (!cond) {
        std::string error_msg;
        if (!msg.empty())
            error_msg = "tgen: " + msg + "\n";
        error_msg += "tgen: THERE IS A BUG IN TGEN; PLEASE CONTACT MAINTAINERS";
        throw std::runtime_error(error_msg);
    }
}
 
// Ensures condition is true, with a clear error message on failure.
#define tgen_ensure(cond, ...)
    if (!(cond))
    tgen::detail::throw_assertion_error(#cond, ##__VA_ARGS__, __FILE__,
                                        __LINE__)
 
// Registering checks.
inline bool registered = false;
inline void ensure_registered() {
    tgen_ensure(registered,
                "tgen was not registered! You should call "
                "tgen::register_gen(argc, argv) before running tgen functions");
}
 
// Template magic to detect types at compile time.
 
// Detects containers != std::string.
template <typename T, typename = void> struct is_container : std::false_type {};
template <typename T>
struct is_container<T,
                    std::void_t<typename std::remove_reference_t<T>::value_type,
                                decltype(std::begin(std::declval<T>())),
                                decltype(std::end(std::declval<T>()))>>
    : std::true_type {};
// Exclude all basic_string variants
template <typename Char, typename Traits, typename Alloc>
struct is_container<std::basic_string<Char, Traits, Alloc>> : std::false_type {
};
template <typename Char, typename Traits, typename Alloc>
struct is_container<const std::basic_string<Char, Traits, Alloc>>
    : std::false_type {};
template <typename Char, typename Traits, typename Alloc>
struct is_container<std::basic_string<Char, Traits, Alloc> &>
    : std::false_type {};
template <typename Char, typename Traits, typename Alloc>
struct is_container<const std::basic_string<Char, Traits, Alloc> &>
    : std::false_type {};
 
// Detects std::pair.
template <typename T> struct is_pair : std::false_type {};
template <typename A, typename B>
struct is_pair<std::pair<A, B>> : std::true_type {};
// Detects std::tuple.
template <typename T> struct is_tuple : std::false_type {};
template <typename... Ts>
struct is_tuple<std::tuple<Ts...>> : std::true_type {};
// Detects scalar (printed atomically).
template <typename T>
struct is_scalar
    : std::bool_constant<!is_container<T>::value and !is_tuple<T>::value and
                         !is_pair<T>::value> {};
// Detects complex container.
template <typename T>
struct is_container_multiline
    : std::bool_constant<is_container<T>::value and
                         !is_scalar<typename std::remove_cv_t<
                             std::remove_reference_t<T>>::value_type>::value> {
};
// Detects complex std::pair.
template <typename T> struct is_pair_multiline : std::false_type {};
template <typename A, typename B>
struct is_pair_multiline<std::pair<A, B>>
    : std::bool_constant<!is_scalar<A>::value or !is_scalar<B>::value> {};
// Detects complex std::tuple.
template <typename Tuple> struct is_tuple_multiline : std::false_type {};
template <typename... Ts>
struct is_tuple_multiline<std::tuple<Ts...>>
    : std::bool_constant<(!is_scalar<Ts>::value or ...)> {};
 
// Used to return false at compile time only if evaluated.
template <typename> inline constexpr bool dependent_false_v = false;
 
/*
 * Properties of custom types.
 */
 
// If type is sequential (list-like).
using is_sequential_tag = void;
 
// Detects associative containers.
template <typename T, typename = void>
struct is_associative_container : std::false_type {};
template <typename T>
struct is_associative_container<
    T, std::void_t<typename T::key_type, typename T::key_compare>>
    : std::true_type {};
 
// Detects sequential generator values.
template <typename T, typename = void>
struct is_sequential : std::false_type {};
template <typename T>
struct is_sequential<
    T, std::void_t<typename std::decay_t<T>::tgen_is_sequential_tag>>
    : std::true_type {};
 
/*
 * Unique rng to use.
 */
 
// The single rng to be used by the library.
inline std::mt19937 rng;
 
/*
 * Printing.
 */
 
// Print view struct for printing either a container or a sequential generator
// element.
template <typename T,
          bool IsCont = detail::is_container<std::decay_t<T>>::value>
struct print_cols_view;
 
// Container.
template <typename T> struct print_cols_view<T, true> {
    const T &value;
    decltype(std::begin(std::declval<const T &>())) it;
 
    print_cols_view(const T &v) : value(v), it(v.begin()) {}
 
    std::size_t size() const { return value.size(); }
    decltype(auto) get(std::size_t) const { return *it; }
    void advance() { ++it; }
};
 
// Sequential generator element.
template <typename T> struct print_cols_view<T, false> {
    const T &value;
 
    print_cols_view(const T &v) : value(v) {}
 
    std::size_t size() const { return value.size(); }
    decltype(auto) get(std::size_t i) const { return value[i]; }
    void advance() {}
};
 
/*
 * Distinct generation.
 */
 
// Rejection cap is multiplier * |seen|; with one value left, falsely reporting
// exhaustion has probability about e^{-84} < 10^{-36}.
constexpr int distinct_attempt_multiplier = 84;
 
// One rejection-sampling step for distinct generation.
// O(T * log k + log^2 k) amortized expected time per call when generating k
// distinct values and next() runs in O(T).
template <typename Seen, typename Fn>
auto try_generate_distinct(Seen &seen, Fn &&next, bool insert = true)
    -> std::optional<std::invoke_result_t<Fn &>> {
    using T = std::invoke_result_t<Fn &>;
    size_t attempts =
        distinct_attempt_multiplier * std::max<size_t>(1, seen.size());
    for (size_t i = 0; i < attempts; ++i) {
        T val = next();
        if (insert) {
            if (seen.insert(val).second)
                return val;
        } else if (seen.count(val) == 0)
            return val;
    }
    return std::nullopt;
}
 
} // namespace detail
 
/*
 * Compiler configuration (see set_compiler).
 */
 
// Kinds of compilers.
enum class compiler_kind { gcc, clang, unknown };
 
// Compiler identity and version.
struct compiler_value {
    compiler_kind kind_;
    int major_;
    int minor_;
 
    compiler_value(compiler_kind kind = compiler_kind::unknown, int major = 0,
                   int minor = 0)
        : kind_(kind), major_(major), minor_(minor) {}
};
 
namespace detail {
 
// Global C++ version value (0 means unknown).
struct cpp_value {
    int version_;
 
    cpp_value(std::optional<int> version = std::nullopt)
        : version_(version ? *version : 0) {
        if (version) {
            tgen_ensure(*version == 17 or *version == 20 or *version == 23,
                        "unsupported C++ version (use 17, 20, 23)");
        }
    }
};
 
inline cpp_value cpp;
inline compiler_value compiler;
 
} // namespace detail
 
/*
 * Base classes.
 */
 
// Needed for return type of some functions.
template <typename T> struct list;
 
// Generates distinct values of a function.
template <typename Func, typename... Args> struct distinct {
    Func func_;
    std::tuple<Args...> args_;
    using T = std::invoke_result_t<Func &, Args &...>;
    std::set<T> seen_;
 
    distinct(Func func, Args... args)
        : func_(std::move(func)), args_(std::move(args)...) {}
 
    // Generates a distinct value (i.e., one not returned before).
    //
    // Assume gen() produces a uniformly random value in O(T) time.
    // Since duplicates are rejected, the expected number of trials over
    // k successful generations is:
    //
    //   sum_{i=1}^k k / i = O(k log k)
    //
    // (coupon collector argument).
    //
    // Each trial additionally performs O(log k) work to check/store
    // previously generated values, yielding a total time of
    // O((T + log k) * k log k).
    //
    // Thus, the amortized expected time per call is
    // O(T * log k + log^2 k).
    //
    // With extremely small probability (< 1e-18), the algorithm may
    // incorrectly report that no more distinct values exist.
    auto gen() {
        auto val = generate_distinct(true);
        if (val)
            return *val;
 
        throw detail::error("distinct: no more distinct values");
    }
    template <typename U> auto gen(std::initializer_list<U> il) {
        return gen(std::vector<U>(il));
    }
 
    // Generates a list of distinct values.
    auto gen_list(int size) {
        std::vector<T> res;
        for (int i = 0; i < size; ++i)
            res.push_back(gen());
 
        return typename list<T>::value(res);
    }
 
    // Checks if there are no more distinct values.
    // With extremely small probability (< 1e-18), the algorithm may
    // incorrectly report that there are no more distinct values.
    bool empty() { return generate_distinct(false) == std::nullopt; }
 
    // Generates all distinct values.
    auto gen_all() {
        std::vector<T> res;
        while (true) {
            auto val = generate_distinct(true);
            if (val)
                res.push_back(*val);
            else
                break;
        }
        return typename list<T>::value(res);
    }
 
    // Nice error for `out << distinct`.
    friend std::ostream &operator<<(std::ostream &out, const distinct &) {
        static_assert(
            detail::dependent_false_v<distinct>,
            "distinct: cannot print a distinct generator. Maybe you forgot to "
            "call `gen()`?");
        return out;
    }
 
  private:
    // Generates distinct value and inserts it if `insert` is true.
    // Returns the value if found, otherwise returns std::nullopt.
    auto generate_distinct(bool insert) {
        return detail::try_generate_distinct(
            seen_, [&] { return std::apply(func_, args_); }, insert);
    }
};
template <typename Func, typename... Args>
distinct(Func, Args...) -> distinct<Func, Args...>;
 
// Base struct for generators.
template <typename Gen> struct gen_base {
    const Gen &self() const { return *static_cast<const Gen *>(this); }
 
    template <typename... Args> auto gen_list(int size, Args &&...args) const {
        std::vector<typename Gen::value> res;
 
        for (int i = 0; i < size; ++i)
            res.push_back(static_cast<const Gen *>(this)->gen(
                std::forward<Args>(args)...));
 
        return typename list<typename Gen::value>::value(res);
    }
 
    // Calls the generator until predicate is true.
    template <typename Pred, typename... Args>
    auto gen_until(Pred predicate, int max_tries, Args &&...args) const {
        for (int i = 0; i < max_tries; ++i) {
            typename Gen::value val = static_cast<const Gen *>(this)->gen(
                std::forward<Args>(args)...);
 
            if (predicate(val))
                return val;
        }
 
        throw detail::error("could not generate value matching predicate");
    }
    template <typename Pred, typename T, typename... Args>
    auto gen_until(Pred predicate, int max_tries, std::initializer_list<T> il,
                   Args &&...args) const {
        return gen_until(predicate, max_tries, std::vector<T>(il),
                         std::forward<Args>(args)...);
    }
 
    // Distinct for generator.
    template <typename... Args> auto distinct(Args &&...args) const {
        return tgen::distinct(
            [self = self()](auto &&...inner_args) mutable -> decltype(auto) {
                return self.gen(
                    std::forward<decltype(inner_args)>(inner_args)...);
            },
            std::forward<Args>(args)...);
    }
    template <typename T, typename... Args>
    auto distinct(std::initializer_list<T> il, Args &&...args) const {
        return distinct(std::vector<T>(il), std::forward<Args>(args)...);
    }
 
    // Nice error for `out << generator`.
    friend std::ostream &operator<<(std::ostream &out, const gen_base &) {
        static_assert(detail::dependent_false_v<gen_base>,
                      "gen_base: cannot print a generator. Maybe you forgot to "
                      "call `gen()`?");
        return out;
    }
};
 
// Base class for generator values.
template <typename Val> struct gen_value_base {
    const Val &self() const { return *static_cast<const Val *>(this); }
 
    bool operator<(const Val &rhs) const {
        return self().to_std() < rhs.to_std();
    }
};
 
namespace detail {
 
// Detects generator values.
template <typename T>
struct is_generator_value
    : std::is_base_of<gen_value_base<std::decay_t<T>>, std::decay_t<T>> {};
 
} // namespace detail
 
/*
 * Easier printing.
 */
 
// Struct to print standard types to std::ostream;
struct print {
    std::string s_;
 
    template <typename T> print(const T &val, char sep = ' ') {
        std::ostringstream oss;
        write(oss, val, sep);
        s_ = oss.str();
    }
    template <typename T>
    print(const std::initializer_list<T> &il, char sep = ' ') {
        std::ostringstream oss;
        write(oss, std::vector<T>(il), sep);
        s_ = oss.str();
    }
    template <typename T>
    print(const std::initializer_list<std::initializer_list<T>> &il,
          char sep = ' ') {
        std::ostringstream oss;
        std::vector<std::vector<T>> mat;
        for (const auto &i : il)
            mat.push_back(i);
        write(oss, mat, sep);
        s_ = oss.str();
    }
 
    template <typename T> void write(std::ostream &os, const T &val, char sep) {
        if constexpr (detail::is_pair<T>::value) {
            if constexpr (detail::is_pair_multiline<T>::value) {
                write(os, val.first, sep);
                os << '\n';
                write(os, val.second, sep);
            } else {
                // Use space for inner separator.
                write(os, val.first, ' ');
                os << sep;
                write(os, val.second, ' ');
            }
        } else if constexpr (detail::is_tuple<T>::value)
            write_tuple(os, val, sep);
        else if constexpr (detail::is_container<T>::value)
            write_container(os, val, sep);
        else if constexpr (std::is_same_v<T, detail::i128> or
                           std::is_same_v<T, detail::u128>)
            write_128_number(os, val);
        else
            os << val;
    }
 
    // Writes 128 bit number.
    template <typename T> void write_128_number(std::ostream &os, T num) {
        static const long long BASE = 1e18;
 
        if (num < 0) {
            os << '-';
            num = -num;
        }
 
        if (num >= BASE) {
            write_128_number(os, num / BASE);
            os << std::setw(18) << std::setfill('0')
               << static_cast<long long>(num % BASE);
        } else
            os << static_cast<long long>(num);
    }
    // Writes container, checking separator.
    template <typename C>
    void write_container(std::ostream &os, const C &container, char sep) {
        bool first = true;
 
        for (const auto &e : container) {
            if (!first)
                os << (detail::is_container_multiline<C>::value ? '\n' : sep);
            first = false;
            write(os, e, detail::is_container_multiline<C>::value ? sep : ' ');
        }
    }
 
    // Writes tuple, checking separator.
    template <typename Tuple, size_t... I>
    void write_tuple_impl(std::ostream &os, const Tuple &tp, char sep,
                          std::index_sequence<I...>) {
        bool first = true;
        ((os << (first ? (first = false, "")
                       : (detail::is_tuple_multiline<Tuple>::value
                              ? "\n"
                              : std::string(1, sep))),
          write(os, std::get<I>(tp),
                detail::is_tuple_multiline<Tuple>::value ? sep : ' ')),
         ...);
    }
    template <typename T>
    void write_tuple(std::ostream &os, const T &tp, char sep) {
        write_tuple_impl(os, tp, sep,
                         std::make_index_sequence<std::tuple_size<T>::value>{});
    }
 
    friend std::ostream &operator<<(std::ostream &out, const print &pr) {
        return out << pr.s_;
    }
};
 
// Prints in a new line.
struct println : print {
    template <typename T>
    println(const T &val, char sep = ' ') : print(val, sep) {}
    template <typename T>
    println(const std::initializer_list<T> &il, char sep = ' ')
        : print(il, sep) {}
    template <typename T>
    println(const std::initializer_list<std::initializer_list<T>> &il,
            char sep = ' ')
        : print(il, sep) {}
 
    friend std::ostream &operator<<(std::ostream &out, const println &pr) {
        return out << pr.s_ << '\n';
    }
};
 
// Prints container / sequential generator value on its own column.
// Example:
//   A = {1, 2, 3}, B = {4, 2, 5}
//   print_each(A, B) will print:
//  "1 4
//   2 2
//   3 5
//",
//  that is, it prints the end of the line for all lines.
template <typename... Args> struct print_cols {
    std::string s_;
 
    print_cols(const Args &...args) {
        static_assert(
            ((detail::is_container<std::decay_t<Args>>::value or
              detail::is_sequential<std::decay_t<Args>>::value) and
             ...),
            "print_cols: arguments must be containers or sequential generator "
            "values");
        std::ostringstream oss;
        write(oss, args...);
        s_ = oss.str();
    }
 
    void write(std::ostream &os, const Args &...args) {
        auto views = std::apply(
            [](const Args &...inner_args) {
                return std::make_tuple(
                    detail::print_cols_view<decltype(inner_args)>{
                        inner_args}...);
            },
            std::forward_as_tuple(args...));
 
        const std::size_t n = std::get<0>(views).size();
 
        auto check = [&](const auto &v) {
            tgen_ensure(v.size() == n, "print_cols: sizes should be the same");
        };
        std::apply([&](const auto &...v) { (check(v), ...); }, views);
 
        for (std::size_t i = 0; i < n; ++i) {
            bool first = true;
 
            std::apply(
                [&](const auto &...v) {
                    ((os << (first ? "" : " ") << print(v.get(i)),
                      first = false),
                     ...);
                },
                views);
 
            os << '\n';
 
            std::apply([](auto &...v) { (v.advance(), ...); }, views);
        }
    }
 
    friend std::ostream &operator<<(std::ostream &out, const print_cols &pr) {
        return out << pr.s_;
    }
};
 
/*
 * Global random operations.
 */
 
namespace detail {
 
// libstdc++ accepts std::uniform_int_distribution with narrow integral types
// (char/signed char/unsigned char/short/bool), but libc++ rejects them with a
// hard static_assert ("IntType must be a supported integer type"). Promote such
// types to a width the standard guarantees, preserving signedness, so the same
// `next<T>` works across both standard libraries (e.g. Apple clang / libc++).
template <typename T>
using uniform_int_t = std::conditional_t<
    (sizeof(T) >= sizeof(short)), T,
    std::conditional_t<std::is_signed_v<T>, int, unsigned int>>;
 
} // namespace detail
 
// Returns a uniformly random number in [0, right)
// O(1).
template <typename T> T next(T right) {
    detail::ensure_registered();
    if constexpr (std::is_integral_v<T>) {
        tgen_ensure(right >= 1, "value for `next` must be valid");
        return static_cast<T>(
            std::uniform_int_distribution<detail::uniform_int_t<T>>(
                0,
                static_cast<detail::uniform_int_t<T>>(right) - 1)(detail::rng));
    } else if constexpr (std::is_floating_point_v<T>) {
        tgen_ensure(right >= 0, "value for `next` must be valid");
        return std::uniform_real_distribution<T>(0, right)(detail::rng);
    } else
        throw detail::error("invalid type for next (" +
                            std::string(typeid(T).name()) + ")");
}
 
// Returns a uniformly random number in [left, right].
// For floating-point types, uses uniform_real_distribution ([left, right) in
// C++), equivalent to [left, right] because the right endpoint has probability
// zero.
// O(1).
template <typename T> T next(T left, T right) {
    detail::ensure_registered();
    tgen_ensure(left <= right, "range for `next` must be valid");
    if constexpr (std::is_integral_v<T>)
        return static_cast<T>(
            std::uniform_int_distribution<detail::uniform_int_t<T>>(
                static_cast<detail::uniform_int_t<T>>(left),
                static_cast<detail::uniform_int_t<T>>(right))(detail::rng));
    else if constexpr (std::is_floating_point_v<T>)
        return std::uniform_real_distribution<T>(left, right)(detail::rng);
    else
        throw detail::error("invalid type for next (" +
                            std::string(typeid(T).name()) + ")");
}
 
// Skewed next.
//
// Returns a random number in [0, right) with a bias controlled by `w`.
// - w = 0:
//     Uniform distribution.
// - w > 0:
//     Returns the maximum of (w + 1) independent uniform samples.
//     Biases the distribution toward larger values.
//     The resulting density is proportional to:
//         f(x) = x^w
//     In particular:
//         w = 1 -> linear bias
//         w = 2 -> quadratic bias
//         w = 3 -> cubic bias
// - w < 0:
//     Returns the minimum of (-w + 1) independent uniform samples.
//     Symmetric to the w > 0 case.
// The continuous version corresponds to Beta distributions:
//     w > 0 -> Beta(w + 1, 1)
//     w < 0 -> Beta(1, -w + 1)
// For |w| > 5, the distribution is approximate.
// O(1).
template <typename T> T wnext(T right, int w) {
    // For small |w|, use the naive approach.
    if (abs(w) <= 5) {
        T val = next<T>(right);
        for (int i = 0; i < w; ++i)
            val = std::max(val, next<T>(right));
        for (int i = 0; i < -w; ++i)
            val = std::min(val, next<T>(right));
        return val;
    }
 
    // O(1) way.
    double x, r = next<double>(0, 1);
 
    if (w >= 0) {
        x = std::pow(r, 1.0 / (w + 1));
    } else {
        x = 1.0 - std::pow(r, 1.0 / (-w + 1));
    }
 
    return T(x * right);
}
 
// Returns a random number in [left, right] with a bias controlled by `w`.
// O(1).
template <typename T> T wnext(T left, T right, int w) {
    // For small |w|, use the naive approach.
    if (abs(w) <= 5) {
        T val = next<T>(left, right);
        for (int i = 0; i < w; ++i)
            val = std::max(val, next<T>(left, right));
        for (int i = 0; i < -w; ++i)
            val = std::min(val, next<T>(left, right));
        return val;
    }
 
    // O(1) way.
    double x, r = next<double>(0, 1);
 
    if (w >= 0) {
        x = std::pow(r, 1.0 / (w + 1));
    } else {
        x = 1.0 - std::pow(r, 1.0 / (-w + 1));
    }
 
    return left + T(x * (right - left));
}
 
namespace detail {
 
// Uniformly random 128 bit number in [0, total).
// O(1) expected.
inline u128 next128(u128 total) {
    tgen_ensure(total > 0, "next128: total must be positive");
 
    // Largest multiple of total less than 2^128.
    u128 limit = (u128(-1) / total) * total;
 
    while (true) {
        // Generate uniform 128-bit random number.
        u128 r = (u128(next<uint64_t>(0, std::numeric_limits<uint64_t>::max()))
                  << 64) |
                 next<uint64_t>(0, std::numeric_limits<uint64_t>::max());
 
        if (r < limit)
            return r % total;
    }
}
 
} // namespace detail
 
// Weighted sampler.
//
// Generates indices with probability proportional to `distribution`, using
// alias method.
//
// Internally, integral weights are accumulated in unsigned __int128 (exact);
// floating-point weights are accumulated in double.
// <O(n), O(1)>.
template <typename T> struct weighted_sampler {
    static_assert(std::is_arithmetic_v<T>,
                  "weighted_sampler requires an arithmetic weight type");
 
    // Internal storage type: `u128` for integral inputs (exact arithmetic),
    // `double` for floating-point inputs.
    using storage_t =
        std::conditional_t<std::is_integral_v<T>, detail::u128, double>;
 
    int n_;
    std::vector<storage_t> weight_;
    std::vector<int> alias_;
    storage_t total_;
 
    // Creates an alias method for generating indices with probability
    // proportional to the distribution.
    // O(n).
    weighted_sampler(const std::vector<T> &distribution)
        : n_(distribution.size()), alias_(n_) {
        tgen_ensure(distribution.size() > 0,
                    "weighted_sampler: distribution must be non-empty");
        for (const auto &w : distribution)
            tgen_ensure(w >= 0,
                        "weighted_sampler: distribution must be non-negative");
 
        total_ = std::accumulate(distribution.begin(), distribution.end(),
                                 storage_t(0));
 
        std::queue<int> big, small;
        for (int i = 0; i < n_; ++i) {
            weight_.push_back(storage_t(n_) * storage_t(distribution[i]));
            if (weight_[i] < total_)
                small.push(i);
            else
                big.push(i);
        }
 
        while (!small.empty() and !big.empty()) {
            int s = small.front();
            small.pop();
            int b = big.front();
            big.pop();
 
            alias_[s] = b;
 
            weight_[b] -= total_ - weight_[s];
            if (weight_[b] < total_)
                small.push(b);
            else
                big.push(b);
        }
 
        detail::tgen_ensure_against_bug(
            small.empty(), "weighted_sampler: small must be empty");
 
        // The remaining elements should have weight equal to total and be
        // assigned to themselves.
        while (!big.empty()) {
            int b = big.front();
            big.pop();
            if constexpr (std::is_integral_v<T>) {
                detail::tgen_ensure_against_bug(
                    weight_[b] == total_,
                    "weighted_sampler: weight of big element must be total");
            }
            alias_[b] = b;
        }
    }
    weighted_sampler(const std::initializer_list<T> &distribution)
        : weighted_sampler(std::vector<T>(distribution)) {}
 
    // Uniformly random value in [0, total). Overloaded so next() can dispatch
    // at compile time to the right primitive for the chosen `storage_t`.
    static detail::u128 sample_below(detail::u128 total) {
        return detail::next128(total);
    }
    static double sample_below(double total) {
        return tgen::next<double>(0, total);
    }
 
    // Generates a random index with probability proportional to the
    // distribution.
    // O(1).
    size_t next() const {
        int i = tgen::next<int>(0, n_ - 1);
        return sample_below(total_) < weight_[i] ? i : alias_[i];
    }
};
template <typename T>
weighted_sampler(const std::vector<T> &) -> weighted_sampler<T>;
template <typename T>
weighted_sampler(const std::initializer_list<T> &) -> weighted_sampler<T>;
 
// Returns i with probability proportional to distribution[i].
// O(|distribution|).
template <typename T>
size_t next_by_distribution(const std::vector<T> &distribution) {
    return weighted_sampler(distribution).next();
}
template <typename T>
size_t next_by_distribution(const std::initializer_list<T> &distribution) {
    return next_by_distribution(std::vector<T>(distribution));
}
 
// Returns a vector of k indices with probability proportional to
// `distribution`. Uses alias method.
// O(k + |distribution|).
template <typename T>
std::vector<int> many_by_distribution(int k,
                                      const std::vector<T> &distribution) {
    tgen_ensure(distribution.size() > 0, "distribution must be non-empty");
    tgen_ensure(k >= 0, "number of elements to choose must be non-negative");
 
    weighted_sampler am(distribution);
    std::vector<int> res;
    for (int i = 0; i < k; ++i)
        res.push_back(am.next());
    return res;
}
template <typename T>
std::vector<int>
many_by_distribution(int k, const std::initializer_list<T> &distribution) {
    return many_by_distribution(k, std::vector<T>(distribution));
}
 
// Shuffles [first, last) inplace uniformly, for RandomAccessIterator.
// O(|container|).
template <typename It> void shuffle(It first, It last) {
    if (first == last)
        return;
 
    for (It i = first + 1; i != last; ++i)
        std::iter_swap(i, first + next(0, static_cast<int>(i - first)));
}
 
// Shuffles container uniformly.
// O(|container|).
template <typename C> [[nodiscard]] auto shuffled(const C &container) {
    if constexpr (detail::is_associative_container<C>::value) {
        std::vector<typename C::value_type> vec(container.begin(),
                                                container.end());
        shuffle(vec.begin(), vec.end());
        return vec;
    } else {
        auto new_container = container;
        shuffle(new_container.begin(), new_container.end());
        return new_container;
    }
}
template <typename T>
[[nodiscard]] std::vector<T> shuffled(const std::initializer_list<T> &il) {
    return shuffled(std::vector<T>(il));
}
 
// Returns a random element from [first, last) uniformly.
// O(1) for random_access_iterator, O(|last - first|) otherwise.
template <typename It> typename It::value_type pick(It first, It last) {
    int size = std::distance(first, last);
    tgen_ensure(size > 0, "cannot pick from empty range");
    It it = first;
    std::advance(it, next(0, size - 1));
    return *it;
}
 
// Returns a random element from container uniformly.
// O(1) for random_access_iterator, O(|container|) otherwise.
template <typename C> typename C::value_type pick(const C &container) {
    return pick(container.begin(), container.end());
}
template <typename T> T pick(const std::initializer_list<T> &il) {
    return pick(std::vector<T>(il));
}
 
// Returns container[i] with probability proportional to distribution[i].
// O(1) for random_access_iterator, O(|container|) otherwise.
template <typename C, typename T>
typename C::value_type pick_by_distribution(const C &container,
                                            std::vector<T> distribution) {
    tgen_ensure(container.size() == distribution.size(),
                "container and distribution must have the same size");
    auto it = container.begin();
    std::advance(it, next_by_distribution(distribution));
    return *it;
}
template <typename C, typename T>
typename C::value_type
pick_by_distribution(const C &container,
                     const std::initializer_list<T> &distribution) {
    return pick_by_distribution(container, std::vector<T>(distribution));
}
template <typename T, typename U>
T pick_by_distribution(const std::initializer_list<T> &il,
                       const std::vector<U> &distribution) {
    return pick_by_distribution(std::vector<T>(il), distribution);
}
template <typename T, typename U>
T pick_by_distribution(const std::initializer_list<T> &il,
                       const std::initializer_list<U> &distribution) {
    return pick_by_distribution(std::vector<T>(il),
                                std::vector<U>(distribution));
}
 
// Chooses k values uniformly from container, as in a subsequence of size k.
// Returns a copy. O(|container|).
template <typename C> C choose(const C &container, int k) {
    tgen_ensure(0 < k and k <= static_cast<int>(container.size()),
                "number of elements to choose must be valid");
    std::vector<typename C::value_type> new_vec;
    C new_container;
    int need = k, left = container.size();
    for (auto cur_it = container.begin(); cur_it != container.end(); ++cur_it) {
        if (next(1, left--) <= need) {
            new_container.insert(new_container.end(), *cur_it);
            need--;
        }
    }
    return new_container;
}
template <typename T>
std::vector<T> choose(const std::initializer_list<T> &il, int k) {
    return choose(std::vector<T>(il), k);
}
 
// Number distinct generator for integral types.
// Optimized for performance (unordered_map virtual list; gen_list uses array
// pool, complement, or sparse sampling).
template <typename T> struct distinct_range {
    T left_, right_;
    T num_available_;
    std::unordered_map<T, T> virtual_list_;
 
    // When the range fits in memory, sample via array Fisher–Yates.
    static constexpr size_t array_pool_max = size_t{1} << 23;
 
    // Generator of distinct values in [left, right].
    distinct_range(T left, T right)
        : left_(left), right_(right), num_available_(right - left + 1) {}
 
    // Returns the number of distinct values left to generate.
    T size() const { return num_available_; }
 
    // Generates a random value in [left_, right_] that has not been generated
    // yet.
    // O(log n).
    T gen() {
        // One iteration of Fisher–Yates.
        tgen_ensure(size() > 0, "distinct_range: no more values to generate");
 
        T i = next<T>(0, size() - 1);
        T j = size() - 1;
 
        auto vi_it = virtual_list_.find(i);
        T vi = vi_it == virtual_list_.end() ? i : vi_it->second;
        auto vj_it = virtual_list_.find(j);
        T vj = vj_it == virtual_list_.end() ? j : vj_it->second;
        virtual_list_[i] = vj;
 
        --num_available_;
 
        return vi + left_;
    }
 
    // Generates a list of distinct values.
    // Optimized for performance (array pool, complement, or sparse sampling).
    // O(size) when the range fits in memory; O(size log range) otherwise.
    auto gen_list(int count) {
        tgen_ensure(count >= 0, "distinct_range: size must be nonnegative");
        tgen_ensure(count <= num_available_,
                    "distinct_range: no more values to generate");
 
        size_t range_size = right_ - left_ + 1;
        size_t sample_count = count;
 
        std::vector<T> res;
        if (sample_count > 0) {
            if (range_size <= array_pool_max)
                res = sample_from_pool(sample_count, range_size);
            else if (sample_count * 2 > range_size)
                res = sample_complement(sample_count, range_size);
            else
                res = sample_sparse(sample_count);
        }
 
        num_available_ -= count;
        virtual_list_.clear();
        return typename list<T>::value(res);
    }
 
    // Generates all distinct values.
    // O(n) when the range fits in memory; O(n log n) otherwise.
    auto gen_all() { return gen_list(size()); }
 
  private:
    // Samples count distinct values via array Fisher–Yates on [left_, right_].
    // O(range_size) time and memory.
    std::vector<T> sample_from_pool(size_t count, size_t range_size) {
        std::vector<T> pool(range_size);
        std::iota(pool.begin(), pool.end(), left_);
        for (size_t i = 0; i < count; ++i) {
            size_t j = next<size_t>(i, range_size - 1);
            std::swap(pool[i], pool[j]);
        }
        pool.resize(count);
        return pool;
    }
 
    // Samples count distinct values by excluding range_size - count values.
    // O(range_size + (range_size - count) log(range_size)).
    std::vector<T> sample_complement(size_t count, size_t range_size) {
        size_t exclude_count = range_size - count;
        std::unordered_set<T> excluded;
        excluded.reserve(exclude_count * 2);
 
        if (exclude_count <= array_pool_max) {
            for (T value : sample_from_pool(exclude_count, range_size))
                excluded.insert(value);
        } else {
            for (T value : sample_sparse(exclude_count))
                excluded.insert(value);
        }
 
        std::vector<T> res;
        res.reserve(count);
        for (T value = left_; value <= right_; ++value) {
            if (!excluded.count(value))
                res.push_back(value);
        }
        detail::tgen_ensure_against_bug(
            res.size() == count, "distinct_range: complement sampling failed");
        return res;
    }
 
    // Samples count distinct values via sparse-map Fisher–Yates.
    // O(count log(range_size)).
    std::vector<T> sample_sparse(size_t count) {
        std::unordered_map<T, T> local_virtual;
        local_virtual.reserve(count * 2);
        T remaining = range_span();
        std::vector<T> res;
        res.reserve(count);
        for (size_t step = 0; step < count; ++step) {
            T i = next<T>(0, remaining - 1);
            T j = remaining - 1;
 
            auto vi_it = local_virtual.find(i);
            T vi = vi_it == local_virtual.end() ? i : vi_it->second;
            auto vj_it = local_virtual.find(j);
            T vj = vj_it == local_virtual.end() ? j : vj_it->second;
            local_virtual[i] = vj;
 
            res.push_back(vi + left_);
            --remaining;
        }
        return res;
    }
 
    // Returns right_ - left_ + 1.
    // O(1).
    T range_span() { return right_ - left_ + 1; }
};
 
// Distinct generator for containers.
template <typename T> struct distinct_container {
    std::vector<T> list_;
    distinct_range<size_t> idx_;
 
    // Creates a distinct container generator for the given container.
    template <typename C>
    distinct_container(const C &container)
        : list_(container.begin(), container.end()),
          idx_(0, static_cast<int>(container.size()) - 1) {}
    distinct_container(const std::initializer_list<T> &il)
        : distinct_container(std::vector<T>(il)) {}
 
    // Returns the number of distinct elements left to generate.
    size_t size() const { return idx_.size(); }
 
    // Generates a random element from container uniformly.
    // O(log n).
    T gen() { return list_[idx_.gen()]; }
 
    // Generates a list of distinct values.
    // O(size * log(n)).
    auto gen_list(int size) {
        std::vector<T> res;
        for (int i = 0; i < size; ++i)
            res.push_back(gen());
        return typename list<T>::value(res);
    }
 
    // Generates all distinct values.
    // O(n log(n))
    auto gen_all() {
        std::vector<T> res;
        while (size() > 0)
            res.push_back(gen());
        return typename list<T>::value(res);
    }
};
template <typename C>
distinct_container(const C &) -> distinct_container<typename C::value_type>;
 
/************
 *          *
 *   OPTS   *
 *          *
 ************/
 
/*
 * Opts - options given to the generator.
 *
 * Incompatible with testlib.
 *
 * Opts are a list of either positional or named options.
 *
 * Named options are given in one of the following formats:
 * 1) -keyname=value or --keyname=value (ex. -n=10   , --test-count=20)
 * 2) -keyname value or --keyname value (ex. -n 10   , --test-count 20)
 *
 * Positional options are numbered from 0 sequentially.
 * For example, for "10 -n=20 str" positional option 1 is the string "str".
 */
 
/*
 * C++ version selection.
 */
 
// Sets C++ version.
inline void set_cpp_version(int version) {
    detail::cpp = detail::cpp_value(version);
}
 
/*
 * Compiler selection.
 */
 
// GCC compiler type.
inline compiler_value gcc(int major = 0, int minor = 0) {
    return {compiler_kind::gcc, major, minor};
}
 
// Clang compiler type.
inline compiler_value clang(int major = 0, int minor = 0) {
    return {compiler_kind::clang, major, minor};
}
 
// Sets compiler.
inline void set_compiler(compiler_value compiler) {
    detail::compiler.kind_ = compiler.kind_;
    detail::compiler.major_ = compiler.major_;
    detail::compiler.minor_ = compiler.minor_;
}
 
namespace detail {
 
// Processes special opt flags.
// Returns true if the key is a special opt flag.
inline bool process_special_opt_flags(std::string &key) {
    // Checks for gen::CPP=17|20|23
    if (key.find("tgen::CPP:") == 0) {
        int prefix_len = std::string("tgen::CPP:").size();
        tgen_ensure(static_cast<int>(key.size()) == prefix_len + 2 and
                        std::isdigit(key[prefix_len]) and
                        std::isdigit(key[prefix_len + 1]),
                    "invalid CPP format");
        int version = std::stoi(key.substr(prefix_len, 2));
        set_cpp_version(version);
        return true;
    }
 
    // Checks for tgen::(GCC|CLANG) or
    // tgen::(GCC|CLANG):(version|version.minor).
    compiler_kind kind;
    size_t prefix_len = 0;
 
    if (key.find("tgen::GCC") == 0) {
        kind = compiler_kind::gcc;
        prefix_len = std::string("tgen::GCC").size();
    } else if (key.find("tgen::CLANG") == 0) {
        kind = compiler_kind::clang;
        prefix_len = std::string("tgen::CLANG").size();
    } else {
        return false;
    }
 
    if (key.size() == prefix_len) {
        set_compiler(compiler_value(kind, 0, 0));
        return true;
    }
 
    tgen_ensure(key[prefix_len] == ':', "invalid compiler format");
    ++prefix_len; // for ':'.
 
    std::string inside = key.substr(prefix_len, key.size() - prefix_len);
    int major = 0, minor = 0;
 
    size_t dot = inside.find('.');
    if (dot == std::string::npos) {
        tgen_ensure(!inside.empty() and
                        std::all_of(inside.begin(), inside.end(), ::isdigit),
                    "invalid compiler version");
        major = std::stoi(inside);
    } else {
        std::string maj = inside.substr(0, dot);
        std::string min = inside.substr(dot + 1);
 
        tgen_ensure(!maj.empty() and
                        std::all_of(maj.begin(), maj.end(), ::isdigit) and
                        maj.size() <= 3,
                    "invalid compiler major version");
        tgen_ensure(!min.empty() and
                        std::all_of(min.begin(), min.end(), ::isdigit) and
                        min.size() <= 3,
                    "invalid compiler minor version");
 
        major = std::stoi(maj);
        minor = std::stoi(min);
    }
 
    set_compiler(compiler_value(kind, major, minor));
 
    return true;
}
 
inline std::vector<std::string>
    pos_opts; // Dictionary containing the positional parsed opts.
inline std::map<std::string, std::string>
    named_opts; // Global dictionary the named parsed opts.
 
template <typename T> T get_opt(const std::string &value) {
    try {
        if constexpr (std::is_same_v<T, bool>) {
            if (value == "true" or value == "1")
                return true;
            if (value == "false" or value == "0")
                return false;
        } else if constexpr (std::is_integral_v<T>) {
            if constexpr (std::is_unsigned_v<T>)
                return static_cast<T>(std::stoull(value));
            else
                return static_cast<T>(std::stoll(value));
        } else if constexpr (std::is_floating_point_v<T>)
            return static_cast<T>(std::stold(value));
        else
            return value; // Default: std::string.
    } catch (...) {
    }
 
    throw error("invalid value `" + value + "` for type " + typeid(T).name());
}
 
inline void parse_opts(int argc, char **argv) {
    // Parses the opts into `pos_opts` vector and `named_opts`
    // map. Starting from 1 to ignore the name of the executable.
    for (int i = 1; i < argc; ++i) {
        std::string key(argv[i]);
 
        if (process_special_opt_flags(key))
            continue;
 
        if (key[0] == '-') {
            tgen_ensure(key.size() > 1,
                        "invalid opt (" + std::string(argv[i]) + ")");
            if ('0' <= key[1] and key[1] <= '9') {
                // This case is a positional negative number argument.
                pos_opts.push_back(key);
                continue;
            }
 
            // Pops first char '-'.
            key = key.substr(1);
        } else {
            // This case is a positional argument that does not start with '-'.
            pos_opts.push_back(key);
            continue;
        }
 
        // Pops a possible second char '-'.
        if (key[0] == '-') {
            tgen_ensure(key.size() > 1,
                        "invalid opt (" + std::string(argv[i]) + ")");
 
            // Pops first char '-'.
            key = key.substr(1);
        }
 
        // Assumes that, if it starts with '-' and second char is not a digit,
        // then it is a <key, value> pair.
        // 1 or 2 chars '-' have already been popped.
 
        std::size_t eq = key.find('=');
        if (eq != std::string::npos) {
            // This is the '--key=value' case.
            std::string value = key.substr(eq + 1);
            key = key.substr(0, eq);
            tgen_ensure(!key.empty() and !value.empty(),
                        "expected non-empty key/value in opt (" +
                            std::string(argv[i]) + ")");
            tgen_ensure(named_opts.count(key) == 0,
                        "cannot have repeated keys");
            named_opts[key] = value;
        } else {
            // This is the '--key value' case.
            tgen_ensure(named_opts.count(key) == 0,
                        "cannot have repeated keys");
            tgen_ensure(argv[i + 1], "value cannot be empty");
            named_opts[key] = std::string(argv[i + 1]);
            ++i;
        }
    }
}
 
inline void set_seed(int argc, char **argv) {
    std::vector<uint32_t> seed;
 
    // Starting from 1 to ignore the name of the executable.
    for (int i = 1; i < argc; ++i) {
        // We append the number of chars, and then the list of chars.
        int size_pos = seed.size();
        seed.push_back(0);
        for (char *s = argv[i]; *s != '\0'; ++s) {
            ++seed[size_pos];
            seed.push_back(*s);
        }
    }
    std::seed_seq seq(seed.begin(), seed.end());
    rng.seed(seq);
}
 
} // namespace detail
 
// Returns true if there is an opt at a given index.
inline bool has_opt(std::size_t index) {
    detail::ensure_registered();
    return index < detail::pos_opts.size();
}
 
// Returns true if there is an opt with a given key.
inline bool has_opt(const std::string &key) {
    detail::ensure_registered();
    return detail::named_opts.count(key) != 0;
}
template <typename K>
std::enable_if_t<std::is_same_v<K, char>, bool> has_opt(K key) {
    return has_opt(std::string(1, key));
}
 
// Returns the parsed opt by a given index. If no opts with the given index are
// found, returns the given default_value.
template <typename T>
T opt(size_t index, std::optional<T> default_value = std::nullopt) {
    detail::ensure_registered();
    if (!has_opt(index)) {
        if (default_value)
            return *default_value;
        throw detail::error("cannot find opt at index " +
                            std::to_string(index));
    }
    return detail::get_opt<T>(detail::pos_opts[index]);
}
 
// Returns the parsed opt by a given key. If no opts with the given key are
// found, returns the given default_value.
template <typename T>
T opt(const std::string &key, std::optional<T> default_value = std::nullopt) {
    detail::ensure_registered();
    if (!has_opt(key)) {
        if (default_value)
            return *default_value;
        throw detail::error("cannot find opt with key " + key);
    }
    return detail::get_opt<T>(detail::named_opts[key]);
}
template <typename T, typename K>
std::enable_if_t<std::is_same_v<K, char>, T>
opt(K key, std::optional<T> default_value = std::nullopt) {
    return opt<T>(std::string(1, key), default_value);
}
 
// Registers generator by initializing rng and parsing opts.
inline void register_gen(int argc, char **argv) {
    detail::set_seed(argc, argv);
 
    detail::pos_opts.clear();
    detail::named_opts.clear();
    detail::parse_opts(argc, argv);
 
    detail::registered = true;
}
 
// Registers generator by initializing rng with a given seed.
inline void register_gen(std::optional<long long> seed = std::nullopt) {
    if (seed)
        detail::rng.seed(*seed);
    else
        detail::rng.seed();
 
    detail::pos_opts.clear();
    detail::named_opts.clear();
 
    detail::registered = true;
}
 
/************
 *          *
 *   LIST   *
 *          *
 ************/
 
/*
 * List generator.
 *
 * List of integral types.
 */
 
template <typename T> struct list : gen_base<list<T>> {
    int size_;            // Size of list.
    T value_l_, value_r_; // Range of defined values.
    std::set<T> values_;  // Set of values. If empty, use range; if not,
                          // represents the possible values, and the range
                          // represents the index in this set.
    std::map<T, int>
        value_idx_in_set_; // Index of every value in the set above.
    mutable std::vector<std::pair<T, T>>
        val_range_; // Range of values of each index.
    mutable std::vector<std::vector<int>> neigh_; // Adjacency list of equality.
    std::vector<std::set<int>>
        diff_restrictions_; // All different restrictions.
    bool index_constraints_{
        false}; // True after fix/equal narrows per-index generation.
    mutable bool uses_full_range_{
        false}; // If true, every index uses [value_l_, value_r_] lazily.
 
    // Creates generator for lists of size 'size', with random T in [value_left,
    // value_right].
    list(int size, T value_left, T value_right)
        : size_(size), value_l_(value_left), value_r_(value_right),
          uses_full_range_(true) {
        tgen_ensure(size_ > 0, "list: size must be positive");
        tgen_ensure(value_l_ <= value_r_, "list: value range must be valid");
    }
 
    // Creates list with value set.
    list(int size, std::set<T> values)
        : size_(size), values_(values), index_constraints_(true) {
        tgen_ensure(size_ > 0, "list: size must be positive");
        tgen_ensure(!values.empty(), "list: value set must be non-empty");
        value_l_ = 0, value_r_ = values.size() - 1;
        val_range_.assign(size_, {value_l_, value_r_});
        int idx = 0;
        for (T val : values_)
            value_idx_in_set_[val] = idx++;
    }
 
    // Restricts lists for list[idx] = val.
    list &fix(int idx, T val) {
        tgen_ensure(0 <= idx and idx < size_, "list: index must be valid");
        ensure_val_range_materialized();
        if (values_.size() == 0) {
            auto &[left, right] = val_range_[idx];
            if (left == right and value_l_ != value_r_) {
                tgen_ensure(left == val,
                            "list: must not set to two different values");
            } else {
                tgen_ensure(left <= val and val <= right,
                            "list: value must be in the defined range");
            }
            left = right = val;
        } else {
            tgen_ensure(values_.count(val),
                        "list: value must be in the set of values");
            auto &[left, right] = val_range_[idx];
            int new_val = value_idx_in_set_[val];
            tgen_ensure(left <= new_val and new_val <= right,
                        "list: must not set to two different values");
            left = right = new_val;
        }
        index_constraints_ = true;
        return *this;
    }
 
    // Restricts lists for list[idx_1] = list[idx_2].
    list &equal(int idx_1, int idx_2) {
        tgen_ensure(0 <= std::min(idx_1, idx_2) and
                        std::max(idx_1, idx_2) < size_,
                    "list: indices must be valid");
        if (idx_1 == idx_2)
            return *this;
 
        ensure_val_range_materialized();
        ensure_neigh_allocated();
        index_constraints_ = true;
        neigh_[idx_1].push_back(idx_2);
        neigh_[idx_2].push_back(idx_1);
        return *this;
    }
 
    // Restricts lists for list[S] to be equal, for given subset S of indices.
    list &equal(std::set<int> indices) {
        if (!indices.empty()) {
            std::set<int>::iterator beg = indices.begin();
            for (auto it = std::next(beg); it != indices.end(); ++it)
                equal(*beg, *it);
        }
        return *this;
    }
 
    // Restricts lists for list[left..right] to have all equal values.
    list &equal_range(int left, int right) {
        tgen_ensure(0 <= left and left <= right and right < size_,
                    "list: range indices must be valid");
        for (int i = left; i < right; ++i)
            equal(i, i + 1);
        return *this;
    }
 
    // Restricts lists for all equal elements.
    list &all_equal() { return equal_range(0, size_ - 1); }
 
    // Restricts lists for list[S] to be different (distinct), for given subset
    // S of indices. You cannot add two of these restrictions on sets that
    // intersect.
    list &different(std::set<int> indices) {
        if (!indices.empty())
            diff_restrictions_.push_back(indices);
        return *this;
    }
 
    // Restricts lists for list[idx_1] != list[idx_2].
    list &different(int idx_1, int idx_2) {
        std::set<int> indices = {idx_1, idx_2};
        return different(indices);
    }
 
    // Restricts lists for list[left..right] to have all different values.
    list &different_range(int left, int right) {
        tgen_ensure(0 <= left and left <= right and right < size_,
                    "list: range indices must be valid");
        std::vector<int> indices(right - left + 1);
        std::iota(indices.begin(), indices.end(), left);
        return different(std::set<int>(indices.begin(), indices.end()));
    }
 
    // Restricts lists for all different elements.
    list &all_different() {
        std::vector<int> indices(size_);
        std::iota(indices.begin(), indices.end(), 0);
        return different(std::set<int>(indices.begin(), indices.end()));
    }
 
    // List value.
    // Operations on a value are not random.
    struct value : gen_value_base<value> {
        using tgen_is_sequential_tag = detail::is_sequential_tag;
 
        using value_type = T;            // Value type, for templates.
        using std_type = std::vector<T>; // std type for value.
 
        std::vector<T> vec_; // list.
        char sep_;           // Separator for printing.
 
        value(const std::vector<T> &vec) : vec_(vec), sep_(' ') {}
        value(const std::initializer_list<T> &il) : value(std::vector<T>(il)) {}
 
        // Fetches size.
        int size() const { return vec_.size(); }
 
        // Fetches position idx.
        T &operator[](int idx) {
            tgen_ensure(0 <= idx and idx < size(),
                        "list: value: index out of bounds");
            return vec_[idx];
        }
        const T &operator[](int idx) const {
            tgen_ensure(0 <= idx and idx < size(),
                        "list: value: index out of bounds");
            return vec_[idx];
        }
 
        // Sorts values in non-decreasing order.
        // O(n log n).
        value &sort() {
            std::sort(vec_.begin(), vec_.end());
            return *this;
        }
 
        // Reverses list.
        // O(n).
        value &reverse() {
            std::reverse(vec_.begin(), vec_.end());
            return *this;
        }
 
        // Sets the separator for the list, for printing.
        // O(1).
        value &separator(char sep) {
            sep_ = sep;
            return *this;
        }
 
        // Concatenates two values.
        // Linear.
        value operator+(const value &rhs) const {
            std::vector<T> new_vec = vec_;
            for (int i = 0; i < rhs.size(); ++i)
                new_vec.push_back(rhs[i]);
            return value(new_vec);
        }
 
        // Shuffles list uniformly.
        // O(n).
        value &shuffle() {
            for (int i = 0; i < size(); ++i)
                std::swap(vec_[i], vec_[next(0, size() - 1)]);
            return *this;
        }
 
        // Returns a random element uniformly.
        // O(1).
        T pick() const { return vec_[next<int>(0, size() - 1)]; }
 
        // Returns vec_[i] with probability proportional to distribution[i].
        // O(1).
        template <typename Dist>
        T pick_by_distribution(const std::vector<Dist> &distribution) const {
            tgen_ensure(static_cast<size_t>(size()) == distribution.size(),
                        "value and distribution must have the same size");
            return vec_[next_by_distribution(distribution)];
        }
        template <typename Dist>
        T pick_by_distribution(
            const std::initializer_list<Dist> &distribution) const {
            return pick_by_distribution(std::vector<Dist>(distribution));
        }
 
        // Chooses k values uniformly, as in a subsequence of size k.
        // O(n).
        value choose(int k) const {
            tgen_ensure(0 < k and k <= size(),
                        "number of elements to choose must be valid");
            std::vector<T> new_vec;
            int need = k;
            for (int i = 0; need > 0; ++i) {
                int left = size() - i;
                if (next(1, left) <= need) {
                    new_vec.push_back(vec_[i]);
                    need--;
                }
            }
            return value(new_vec);
        }
 
        // Prints to std::ostream, separated by sep_.
        friend std::ostream &operator<<(std::ostream &out, const value &val) {
            for (int i = 0; i < val.size(); ++i) {
                if (i > 0)
                    out << val.sep_;
                out << val[i];
            }
            return out;
        }
 
        // Gets a std::vector representing the value.
        auto to_std() const {
            if constexpr (!detail::is_generator_value<T>::value) {
                return vec_;
            } else {
                std::vector<typename T::std_type> vec;
                for (const auto &i : vec_)
                    vec.push_back(i.to_std());
                return vec;
            }
        }
    };
 
    // Generates list value.
    // Optimized for performance (unconstrained and all-different fast paths).
    // O(n log n).
    value gen() const {
        if (diff_restrictions_.empty()) {
            if (auto unconstrained = try_gen_unconstrained())
                return *unconstrained;
        }
        if (auto all_different = try_gen_all_different())
            return *all_different;
 
        ensure_neigh_allocated();
        std::vector<T> vec(size_);
        std::vector<bool> defined_idx(
            size_, false); // For every index, if it has been set in `vec`.
 
        std::vector<int> comp_id(size_, -1); // Component id of each index.
        std::vector<std::vector<int>> comp(size_); // Component of each comp-id.
        int comp_count = 0; // Number of different components.
 
        // Defines value of entire component.
        auto define_comp = [&](int cur_comp, T val) {
            for (int idx : comp[cur_comp]) {
                tgen_ensure(!defined_idx[idx]);
                vec[idx] = val;
                defined_idx[idx] = true;
            }
        };
 
        // Groups = components.
        {
            std::vector<bool> vis(size_, false); // Visited for each index.
            for (int idx = 0; idx < size_; ++idx)
                if (!vis[idx]) {
                    T new_value;
                    bool value_defined = false;
 
                    // BFS to visit the connected component, grouping equal
                    // values.
                    std::queue<int> q({idx});
                    vis[idx] = true;
                    std::vector<int> component;
                    while (!q.empty()) {
                        int cur_idx = q.front();
                        q.pop();
 
                        component.push_back(cur_idx);
 
                        // Checks value.
                        auto [l, r] = val_range_at(cur_idx);
                        if (l == r) {
                            if (!value_defined) {
                                // We found the value.
                                value_defined = true;
                                new_value = l;
                            } else if (new_value != l) {
                                // We found a contradiction
                                throw detail::contradiction_error(
                                    "list",
                                    "tried to set value to `" +
                                        std::to_string(new_value) +
                                        "`, but it was already set as `" +
                                        std::to_string(l) + "`");
                            }
                        }
 
                        for (int nxt_idx : neigh_[cur_idx]) {
                            if (!vis[nxt_idx]) {
                                vis[nxt_idx] = true;
                                q.push(nxt_idx);
                            }
                        }
                    }
 
                    // Group entire component, checking if value is defined.
                    for (int cur_idx : component) {
                        comp_id[cur_idx] = comp_count;
                        comp[comp_id[cur_idx]].push_back(cur_idx);
                    }
 
                    // Defines value if needed.
                    if (value_defined)
                        define_comp(comp_count, new_value);
 
                    ++comp_count;
                }
        }
 
        // Initial parsing of different restrictions.
        std::vector<std::set<int>> diff_containing_comp_idx(comp_count);
        {
            int dist_id = 0;
            for (const std::set<int> &diff : diff_restrictions_) {
                // Checks if there are too many different values.
                if (static_cast<uint64_t>(diff.size() - 1) +
                        static_cast<uint64_t>(value_l_) >
                    static_cast<uint64_t>(value_r_))
                    throw detail::contradiction_error(
                        "list", "tried to generate " +
                                    std::to_string(diff.size()) +
                                    " different values, but the maximum is " +
                                    std::to_string(value_r_ - value_l_ + 1));
 
                // Checks if two values in same component are marked as
                // different.
                std::set<int> comp_ids;
                for (int idx : diff) {
                    if (comp_ids.count(comp_id[idx]))
                        throw detail::contradiction_error(
                            "list", "tried to set two indices as equal and "
                                    "different");
                    comp_ids.insert(comp_id[idx]);
 
                    diff_containing_comp_idx[comp_id[idx]].insert(dist_id);
                }
                ++dist_id;
            }
        }
 
        // If some value is in >= 3 sets, then there is a cycle.
        for (auto &diff_containing : diff_containing_comp_idx)
            if (diff_containing.size() >= 3)
                throw detail::complex_restrictions_error(
                    "list",
                    "one index cannot be in >= 3 'different' restrictions");
 
        std::vector<bool> vis_diff(diff_restrictions_.size(), false);
        std::vector<bool> initially_defined_comp_idx(comp_count, false);
 
        // Fills the value in a tree defined by "different" restrictions.
        auto define_tree = [&](int diff_id) {
            // The set `diff_restrictions_[diff_id]` can have some
            // values that are defined.
 
            // Generates set of already defined values.
            std::set<T> defined_values;
            for (int idx : diff_restrictions_[diff_id])
                if (defined_idx[idx]) {
                    // Checks if two values in `diff_restrictions_[dist_id]`
                    // have been set to the same value
                    if (defined_values.count(vec[idx]))
                        throw detail::contradiction_error(
                            "list",
                            "tried to set two indices as equal and different");
 
                    defined_values.insert(vec[idx]);
                }
 
            // Generates values in this root "different" restriction.
            {
                int new_value_count = diff_restrictions_[diff_id].size() -
                                      static_cast<int>(defined_values.size());
                std::vector<T> generated_values =
                    generate_distinct_values(new_value_count, defined_values);
                auto val_it = generated_values.begin();
                for (int idx : diff_restrictions_[diff_id])
                    if (defined_idx[idx]) {
                        // The root can cover these components, but there should
                        // not be any other defined in this tree.
                        initially_defined_comp_idx[comp_id[idx]] = false;
                    } else {
                        define_comp(comp_id[idx], *val_it);
                        ++val_it;
                    }
            }
 
            // BFS on the tree of "different" restrictions.
            std::queue<std::pair<int, int>> q; // {id, parent id}
            q.emplace(diff_id, -1);
            vis_diff[diff_id] = true;
            while (!q.empty()) {
                auto [cur_diff, parent] = q.front();
                q.pop();
 
                std::set<int> neigh_diff;
                for (int idx : diff_restrictions_[cur_diff])
                    for (int nxt_diff :
                         diff_containing_comp_idx[comp_id[idx]]) {
                        if (nxt_diff == cur_diff or nxt_diff == parent)
                            continue;
 
                        // Cycle found.
                        if (vis_diff[nxt_diff])
                            throw detail::complex_restrictions_error(
                                "list",
                                "cycle found in 'different' restrictions");
 
                        neigh_diff.insert(nxt_diff);
                    }
 
                for (int nxt_diff : neigh_diff) {
                    vis_diff[nxt_diff] = true;
                    q.emplace(nxt_diff, cur_diff);
 
                    // Generates this "different" restriction.
                    std::set<T> nxt_defined_values;
                    for (int idx2 : diff_restrictions_[nxt_diff])
                        if (defined_idx[idx2]) {
                            // There cannot be any more defined. This case is
                            // when there are values not covered by a single
                            // "different" restriction in the tree.
                            if (initially_defined_comp_idx[comp_id[idx2]])
                                throw detail::complex_restrictions_error(
                                    "list");
 
                            nxt_defined_values.insert(vec[idx2]);
                        }
                    int new_value_count =
                        diff_restrictions_[nxt_diff].size() -
                        static_cast<int>(nxt_defined_values.size());
                    std::vector<T> generated_values = generate_distinct_values(
                        new_value_count, nxt_defined_values);
                    auto val_it = generated_values.begin();
                    for (int idx2 : diff_restrictions_[nxt_diff])
                        if (!defined_idx[idx2]) {
                            define_comp(comp_id[idx2], *val_it);
                            ++val_it;
                        }
                }
            }
        };
 
        // Loops through "different" restrictions, sorts "different"
        // restrictions by number of defined components (non-increasing). This
        // guarantees that if there is a valid root (that covers all 'defined'),
        // we find it.
        {
            std::vector<std::pair<int, int>> defined_cnt_and_diff_idx;
            int dist_id = 0;
            for (const std::set<int> &diff : diff_restrictions_) {
                int defined_cnt = 0;
                for (int idx : diff)
                    if (defined_idx[idx]) {
                        ++defined_cnt;
                        initially_defined_comp_idx[comp_id[idx]] = true;
                    }
                defined_cnt_and_diff_idx.emplace_back(defined_cnt, dist_id);
                ++dist_id;
            }
 
            std::sort(defined_cnt_and_diff_idx.rbegin(),
                      defined_cnt_and_diff_idx.rend());
            for (auto [defined_cnt, diff_idx] : defined_cnt_and_diff_idx)
                if (!vis_diff[diff_idx])
                    define_tree(diff_idx);
        }
 
        // Loops through "different" restrictions do define the rest.
        for (std::size_t dist_id = 0; dist_id < diff_restrictions_.size();
             ++dist_id)
            if (!vis_diff[dist_id])
                define_tree(dist_id);
 
        // Define final values. These values all should be random in [l, r], and
        // the "different" restrictions have already been processed. However,
        // there can be still equality restrictions, so we define entire
        // components.
        for (int idx = 0; idx < size_; ++idx)
            if (!defined_idx[idx])
                define_comp(comp_id[idx], next<T>(value_l_, value_r_));
 
        if (!values_.empty()) {
            // Needs to fetch the values from the value set.
            std::vector<T> value_vec(values_.begin(), values_.end());
            for (T &val : vec)
                val = value_vec[val];
        }
 
        return value(vec);
    }
 
  private:
    // Materializes neigh_ after the first equality restriction.
    void ensure_neigh_allocated() const {
        if (neigh_.size() == static_cast<size_t>(size_))
            return;
        neigh_.assign(size_, {});
    }
 
    // Materializes val_range_ after the first per-index restriction.
    void ensure_val_range_materialized() const {
        if (!uses_full_range_)
            return;
        val_range_.assign(size_, {value_l_, value_r_});
        uses_full_range_ = false;
    }
 
    // Returns the allowed value range at index idx.
    std::pair<T, T> val_range_at(int idx) const {
        if (uses_full_range_)
            return {value_l_, value_r_};
        return val_range_[idx];
    }
 
    // Generates a uniformly random list of k distinct values in `[value_l,
    // value_r]`, such that no value is in `forbidden_values`.
    std::vector<T>
    generate_distinct_values(int k, const std::set<T> &forbidden_values) const {
        for (auto forbidden : forbidden_values)
            tgen_ensure(value_l_ <= forbidden and forbidden <= value_r_);
        const T num_available =
            (value_r_ - value_l_ + 1) - forbidden_values.size();
        if (num_available < k)
            throw detail::complex_restrictions_error(
                "list", "not enough distinct values");
        if (forbidden_values.empty())
            return distinct_range<T>(value_l_, value_r_).gen_list(k).to_std();
 
        std::map<T, T> virtual_list;
        std::vector<T> gen_list;
        for (int i = 0; i < k; ++i) {
            T j = next<T>(i, num_available - 1);
            T vj = virtual_list.count(j) ? virtual_list[j] : j;
            T vi = virtual_list.count(i) ? virtual_list[i] : i;
 
            virtual_list[j] = vi, virtual_list[i] = vj;
 
            gen_list.push_back(virtual_list[i]);
        }
 
        for (T &val : gen_list)
            val += value_l_;
 
        std::vector<std::pair<T, int>> values_sorted;
        for (std::size_t i = 0; i < gen_list.size(); ++i)
            values_sorted.emplace_back(gen_list[i], i);
        std::sort(values_sorted.begin(), values_sorted.end());
        auto cur_it = forbidden_values.begin();
        int smaller_forbidden_count = 0;
        for (auto [val, idx] : values_sorted) {
            while (cur_it != forbidden_values.end() and
                   *cur_it <= val + smaller_forbidden_count)
                ++cur_it, ++smaller_forbidden_count;
            gen_list[idx] += smaller_forbidden_count;
        }
 
        return gen_list;
    }
 
    // If this generator has no constraints beyond [value_l_, value_r_],
    // returns independent uniform samples; otherwise returns std::nullopt.
    // O(n).
    std::optional<value> try_gen_unconstrained() const {
        if (!values_.empty() or index_constraints_)
            return std::nullopt;
 
        std::vector<T> vec(size_);
        for (int i = 0; i < size_; ++i)
            vec[i] = next<T>(value_l_, value_r_);
        return value(vec);
    }
 
    // If this generator is exactly all-distinct in [value_l_, value_r_],
    // returns a uniformly random list; otherwise returns std::nullopt.
    // Optimized for performance (distinct_range fast path).
    // O(n log n).
    std::optional<value> try_gen_all_different() const {
        if (!values_.empty() or diff_restrictions_.size() != 1)
            return std::nullopt;
 
        const std::set<int> &diff = diff_restrictions_[0];
        if (static_cast<int>(diff.size()) != size_ or *diff.begin() != 0 or
            *diff.rbegin() != size_ - 1)
            return std::nullopt;
 
        if (!neigh_.empty()) {
            for (const auto &adj : neigh_) {
                if (!adj.empty())
                    return std::nullopt;
            }
        }
 
        if (index_constraints_)
            return std::nullopt;
 
        if (static_cast<long long>(size_) >
            static_cast<long long>(value_r_) - value_l_ + 1)
            throw detail::contradiction_error(
                "list", "tried to generate " + std::to_string(size_) +
                            " different values, but the maximum is " +
                            std::to_string(value_r_ - value_l_ + 1));
 
        return distinct_range<T>(value_l_, value_r_).gen_list(size_);
    }
};
 
/*******************
 *                 *
 *   PERMUTATION   *
 *                 *
 *******************/
 
/*
 * Permutation generation.
 *
 * Permutation are defined always as numbers in [0, n), that is, 0-based.
 */
 
struct permutation : gen_base<permutation> {
    int size_;                                    // Size of permutation.
    std::vector<std::pair<int, int>> defs_;       // {idx, value}.
    std::optional<std::vector<int>> cycle_sizes_; // Cycle sizes.
 
    // Creates generator for permutation of size 'size'.
    permutation(int size) : size_(size) {
        tgen_ensure(size_ > 0, "permutation: size must be positive");
    }
 
    // Restricts permutations for permutation[idx] = val.
    permutation &fix(int idx, int val) {
        tgen_ensure(0 <= idx and idx < size_,
                    "permutation: index must be valid");
        defs_.emplace_back(idx, val);
        return *this;
    }
 
    // Restricts permutations for permutation to have cycle sizes.
    permutation &cycles(const std::vector<int> &cycle_sizes) {
        tgen_ensure(
            size_ == std::accumulate(cycle_sizes.begin(), cycle_sizes.end(), 0),
            "permutation: cycle sizes must add up to size of permutation");
        cycle_sizes_ = cycle_sizes;
        return *this;
    }
    permutation &cycles(const std::initializer_list<int> &cycle_sizes) {
        return cycles(std::vector<int>(cycle_sizes));
    }
 
    // Permutation value.
    // Operations on a value are not random.
    struct value : gen_value_base<value> {
        using tgen_is_sequential_tag = detail::is_sequential_tag;
 
        using std_type = std::vector<int>; // std type for value.
        std::vector<int> vec_;             // Permutation.
        char sep_;                         // Separator for printing.
        bool add_1_;                       // If should add 1, for printing.
 
        value(const std::vector<int> &vec)
            : vec_(vec), sep_(' '), add_1_(false) {
            tgen_ensure(!vec_.empty(), "permutation: value: cannot be empty");
            std::vector<bool> vis(vec_.size(), false);
            for (int i = 0; i < size(); ++i) {
                tgen_ensure(0 <= vec_[i] and
                                vec_[i] < static_cast<int>(vec_.size()),
                            "permutation: value: values must be from `0` to "
                            "`size-1`");
                tgen_ensure(!vis[vec_[i]],
                            "permutation: value: cannot have repeated values");
                vis[vec_[i]] = true;
            }
        }
        value(const std::initializer_list<int> &il)
            : value(std::vector<int>(il)) {}
 
        // Fetches size.
        int size() const { return vec_.size(); }
 
        // Fetches position idx.
        const int &operator[](int idx) const {
            tgen_ensure(0 <= idx and idx < size(),
                        "permutation: value: index out of bounds");
            return vec_[idx];
        }
 
        // Returns parity of the permutation (+1 if even, -1 if odd).
        // O(n).
        int parity() const {
            std::vector<bool> vis(size(), false);
            int cycles = 0;
 
            for (int i = 0; i < size(); ++i)
                if (!vis[i]) {
                    ++cycles;
                    for (int j = i; !vis[j]; j = vec_[j])
                        vis[j] = true;
                }
            // Even iff (n - cycles) is even.
            return ((size() - cycles) % 2 == 0) ? +1 : -1;
        }
 
        // Sorts values in increasing order.
        // O(n).
        value &sort() {
            for (int i = 0; i < size(); ++i)
                vec_[i] = i;
            return *this;
        }
 
        // Reverses permutation.
        // O(n).
        value &reverse() {
            std::reverse(vec_.begin(), vec_.end());
            return *this;
        }
 
        // Inverse of the permutation.
        // O(n).
        value &inverse() {
            std::vector<int> inv(size());
            for (int i = 0; i < size(); ++i)
                inv[vec_[i]] = i;
            swap(vec_, inv);
            return *this;
        }
 
        // Sets the separator, for printing.
        // O(1).
        value &separator(char sep) {
            sep_ = sep;
            return *this;
        }
 
        // Sets that should print values 1-based.
        // O(1).
        value &add_1() {
            add_1_ = true;
            return *this;
        }
 
        // Shuffles permutation uniformly.
        // O(n).
        value &shuffle() {
            for (int i = 0; i < size(); ++i)
                std::swap(vec_[i], vec_[next(0, size() - 1)]);
            return *this;
        }
 
        // Returns a random element uniformly.
        // O(1).
        int pick() const { return vec_[next<int>(0, size() - 1)]; }
 
        // Returns vec_[i] with probability proportional to distribution[i].
        // O(1).
        template <typename Dist>
        int pick_by_distribution(const std::vector<Dist> &distribution) const {
            tgen_ensure(static_cast<size_t>(size()) == distribution.size(),
                        "value and distribution must have the same size");
            return vec_[next_by_distribution(distribution)];
        }
        template <typename Dist>
        int pick_by_distribution(
            const std::initializer_list<Dist> &distribution) const {
            return pick_by_distribution(std::vector<Dist>(distribution));
        }
 
        // Prints to std::ostream, separated by sep_.
        friend std::ostream &operator<<(std::ostream &out, const value &val) {
            for (int i = 0; i < val.size(); ++i) {
                if (i > 0)
                    out << val.sep_;
                out << val[i] + val.add_1_;
            }
            return out;
        }
 
        // Gets a std::vector representing the value.
        std::vector<int> to_std() const { return std_type(vec_); }
    };
 
    // Generates permutation value.
    // O(n).
    value gen() const {
        if (!cycle_sizes_) {
            // Cycle sizes not specified.
            std::vector<int> idx_to_val(size_, -1), val_to_idx(size_, -1);
            for (auto [idx, val] : defs_) {
                tgen_ensure(
                    0 <= val and val < size_,
                    "permutation: value in permutation must be in [0, " +
                        std::to_string(size_) + ")");
 
                if (idx_to_val[idx] != -1) {
                    tgen_ensure(idx_to_val[idx] == val,
                                "permutation: cannot set an index to two "
                                "different values");
                } else
                    idx_to_val[idx] = val;
 
                if (val_to_idx[val] != -1) {
                    tgen_ensure(val_to_idx[val] == idx,
                                "permutation: cannot set two indices to the "
                                "same value");
                } else
                    val_to_idx[val] = idx;
            }
 
            std::vector<int> perm(size_);
            std::iota(perm.begin(), perm.end(), 0);
            shuffle(perm.begin(), perm.end());
            int cur_idx = 0;
            for (int &i : idx_to_val)
                if (i == -1) {
                    // While this value is used, skip.
                    while (val_to_idx[perm[cur_idx]] != -1)
                        ++cur_idx;
                    i = perm[cur_idx++];
                }
            return idx_to_val;
        }
 
        // Creates cycles.
        std::vector<int> order(size_);
        std::iota(order.begin(), order.end(), 0);
        shuffle(order.begin(), order.end());
        int idx = 0;
        std::vector<std::vector<int>> cycles;
        for (int cycle_size : *cycle_sizes_) {
            cycles.emplace_back();
            for (int i = 0; i < cycle_size; ++i)
                cycles.back().push_back(order[idx++]);
        }
 
        // Retrieves permutation from cycles.
        std::vector<int> perm(size_, -1);
        for (const std::vector<int> &cycle : cycles) {
            int cur_size = cycle.size();
            for (int i = 0; i < cur_size; ++i)
                perm[cycle[i]] = cycle[(i + 1) % cur_size];
        }
 
        return value(perm);
    }
};
 
/************
 *          *
 *   MATH   *
 *          *
 ************/
 
namespace math {
 
namespace detail {
 
using namespace tgen::detail;
 
inline int popcount(uint64_t x) { return __builtin_popcountll(x); }
 
inline int ctzll(uint64_t x) {
    // Mystery code found on the internet.
    // Uses de Bruijn sequence.
    static const unsigned char index64[64] = {
        0,  1,  2,  53, 3,  7,  54, 27, 4,  38, 41, 8,  34, 55, 48, 28,
        62, 5,  39, 46, 44, 42, 22, 9,  24, 35, 59, 56, 49, 18, 29, 11,
        63, 52, 6,  26, 37, 40, 33, 47, 61, 45, 43, 21, 23, 58, 17, 10,
        51, 25, 36, 32, 60, 20, 57, 16, 50, 31, 19, 15, 30, 14, 13, 12};
    return index64[((x & -x) * 0x022FDD63CC95386D) >> 58];
}
 
inline uint64_t mul_mod(uint64_t a, uint64_t b, uint64_t m) {
    return static_cast<u128>(a) * b % m;
}
 
// O(log n).
// 0 <= x < m.
inline uint64_t expo_mod(uint64_t x, uint64_t y, uint64_t m) {
    if (!y)
        return 1;
    uint64_t ans = expo_mod(mul_mod(x, x, m), y / 2, m);
    return y % 2 ? mul_mod(x, ans, m) : ans;
}
 
} // namespace detail
 
// O(log^2 n).
inline bool is_prime(uint64_t n) {
    if (n < 2)
        return false;
    if (n == 2 or n == 3)
        return true;
    if (n % 2 == 0)
        return false;
 
    uint64_t r = detail::ctzll(n - 1), d = n >> r;
    // These bases are guaranteed to work for n <= 2^64.
    for (int a : {2, 325, 9375, 28178, 450775, 9780504, 1795265022}) {
        uint64_t x = detail::expo_mod(a, d, n);
        if (x == 1 or x == n - 1 or a % n == 0)
            continue;
 
        for (uint64_t j = 0; j < r - 1; ++j) {
            x = detail::mul_mod(x, x, n);
            if (x == n - 1)
                break;
        }
        if (x != n - 1)
            return false;
    }
    return true;
}
 
namespace detail {
 
inline uint64_t pollard_rho(uint64_t n) {
    if (n == 1 or is_prime(n))
        return n;
    auto f = [n](uint64_t x) { return mul_mod(x, x, n) + 1; };
 
    uint64_t x = 0, y = 0, t = 30, prd = 2, x0 = 1, q;
    while (t % 40 != 0 or std::gcd(prd, n) == 1) {
        if (x == y)
            x = ++x0, y = f(x);
        q = mul_mod(prd, x > y ? x - y : y - x, n);
        if (q != 0)
            prd = q;
        x = f(x), y = f(f(y)), ++t;
    }
    return std::gcd(prd, n);
}
 
inline std::vector<uint64_t> factor(uint64_t n) {
    if (n == 1)
        return {};
    if (is_prime(n))
        return {n};
    uint64_t d = pollard_rho(n);
    std::vector<uint64_t> l = factor(d), r = factor(n / d);
    l.insert(l.end(), r.begin(), r.end());
    return l;
}
 
// Error handling.
template <typename T>
std::runtime_error there_is_no_in_range_error(const std::string &type, T l,
                                              T r) {
    return error("math: there is no " + type + " in range [" +
                 std::to_string(l) + ", " + std::to_string(r) + "]");
}
template <typename T>
std::runtime_error there_is_no_from_error(const std::string &type, T r) {
    return error("math: there is no " + type + " from " + std::to_string(r));
}
template <typename T>
std::runtime_error there_is_no_upto_error(const std::string &type, T r) {
    return error("math: there is no " + type + " up to " + std::to_string(r));
}
 
// O(log mod).
// 0 < a < mod.
// gcd(a, mod) = 1.
inline i128 modular_inverse_128(i128 a, i128 mod) {
    tgen_ensure(0 < a and a < mod,
                "math: modular inverse requires 0 < value < mod");
 
    i128 t = 0, new_t = 1;
    i128 r = mod, new_r = a;
 
    while (new_r != 0) {
        i128 q = r / new_r;
 
        auto tmp_t = t - q * new_t;
        t = new_t;
        new_t = tmp_t;
 
        auto tmp_r = r - q * new_r;
        r = new_r;
        new_r = tmp_r;
    }
 
    tgen_ensure(r == 1, "math: remainder and mod must be coprime");
 
    if (t < 0)
        t += mod;
    return t;
}
 
// checks if a * b <= limit, for positive numbers.
inline bool mul_leq(uint64_t a, uint64_t b, uint64_t limit) {
    if (a == 0 or b == 0)
        return true;
    return a <= limit / b;
}
 
// base^exp, or null if base^exp > limit.
inline std::optional<uint64_t> expo(uint64_t base, uint64_t exp,
                                    uint64_t limit) {
    uint64_t result = 1;
 
    while (exp) {
        if (exp & 1) {
            if (!mul_leq(result, base, limit))
                return std::nullopt;
            result *= base;
        }
 
        exp >>= 1;
        // Necessary for correctness.
        if (!exp)
            break;
 
        if (!mul_leq(base, base, limit))
            return std::nullopt;
        base *= base;
    }
    return result;
}
 
// O(log n log k).
// 0 < k.
inline uint64_t kth_root_floor(uint64_t n, uint64_t k) {
    tgen_ensure_against_bug(k > 0, "math: value must be valid");
    if (k == 1 or n <= 1)
        return n;
 
    uint64_t lo = 1, hi = 1ULL << ((64 + k - 1) / k);
 
    while (lo < hi) {
        uint64_t mid = lo + (hi - lo + 1) / 2;
 
        if (expo(mid, k, n)) {
            lo = mid;
        } else {
            hi = mid - 1;
        }
    }
    return lo;
}
 
// gcd(a, b).
// O(log a).
inline i128 gcd128(i128 a, i128 b) {
    if (a < 0)
        a = -a;
    if (b < 0)
        b = -b;
    while (b != 0) {
        i128 t = a % b;
        a = b;
        b = t;
    }
    return a;
}
 
// min(2^64, a*b).
// O(log a).
// a, b >= 0.
inline i128 mul_saturate(i128 a, i128 b) {
    tgen_ensure(a >= 0 and b >= 0);
    static const i128 LIMIT = static_cast<i128>(1) << 64;
    if (a == 0 or b == 0)
        return 0;
    if (a > LIMIT / b)
        return LIMIT;
    return a * b;
}
 
struct crt {
    using T = i128;
    T a, m;
 
    crt() : a(0), m(1) {}
    crt(T a_, T m_) : a(a_), m(m_) {}
    crt operator*(crt C) {
        if (m == 0 or C.m == 0)
            return {-1, 0};
 
        T g = gcd128(m, C.m);
        if ((C.a - a) % g != 0)
            return {-1, 0};
 
        T m1 = m / g;
        T m2 = C.m / g;
 
        if (m2 == 1)
            return {a, m};
 
        T inv = modular_inverse_128(m1 % m2, m2);
 
        T k = ((C.a - a) / g) % m2;
        if (k < 0)
            k += m2;
 
        k = static_cast<u128>(k) * inv % m2;
 
        T lcm = mul_saturate(m, m2);
 
        T res = (a + static_cast<T>((static_cast<u128>(k) * m) % lcm)) % lcm;
        if (res < 0)
            res += lcm;
 
        return {res, lcm};
    }
};
 
// Math hacks to operate on log space.
 
inline constexpr long double LOG_ZERO = -INFINITY;
inline constexpr long double LOG_ONE = 0.0;
 
inline long double log_space(long double x) {
    return x == 0.0 ? LOG_ZERO : std::log(x);
}
 
// Math hack to add two values in log space.
inline long double add_log_space(long double a, long double b) {
    if (a < b)
        std::swap(a, b);
    if (b == LOG_ZERO)
        return a;
    return a + log1p(exp(b - a));
}
 
// Math hack to subtract two values in log space.
// a >= b.
inline long double sub_log_space(long double a, long double b) {
    if (b >= a)
        return LOG_ZERO;
    if (b == LOG_ZERO)
        return a;
    return a + log1p(-exp(b - a));
}
 
} // namespace detail
 
// Sorted.
// O(n^(1/4) log n) expected.
// 0 < n.
inline std::vector<uint64_t> factor(uint64_t n) {
    tgen_ensure(n > 0, "math: number to factor must be positive");
    auto factors = detail::factor(n);
    std::sort(factors.begin(), factors.end());
    return factors;
}
 
// Sorted.
// O(n^(1/4) log n) expected.
// 0 < n.
inline std::vector<std::pair<uint64_t, int>> factor_by_prime(uint64_t n) {
    tgen_ensure(n > 0, "math: number to factor must be positive");
    std::vector<std::pair<uint64_t, int>> primes;
    for (uint64_t p : factor(n)) {
        if (!primes.empty() and primes.back().first == p)
            ++primes.back().second;
        else
            primes.emplace_back(p, 1);
    }
    return primes;
}
 
// O(log mod).
// 0 < a < mod.
// gcd(a, mod) = 1.
inline uint64_t modular_inverse(uint64_t a, uint64_t mod) {
    return detail::modular_inverse_128(a, mod);
}
 
// O(n^(1/4) log n) expected.
// 0 < n.
inline uint64_t totient(uint64_t n) {
    tgen_ensure(n > 0, "math: totient(0) is undefined");
    uint64_t phi = n;
 
    for (auto [p, e] : factor_by_prime(n))
        phi -= phi / p;
 
    return phi;
}
 
// Returns `(p_i, g_i)`: `p_i` is the prime, `g_i` is the gap.
inline const std::pair<std::vector<uint64_t>, std::vector<uint64_t>> &
prime_gaps() {
    // From https://en.wikipedia.org/wiki/Prime_gap.
    static const std::pair<std::vector<uint64_t>, std::vector<uint64_t>> value{
        /* clang-format off */ {
            2, 3, 7, 23, 89, 113, 523, 887, 1129, 1327, 9551, 15683, 19609,
            31397, 155921, 360653, 370261, 492113, 1349533, 1357201, 2010733,
            4652353, 17051707, 20831323, 47326693, 122164747, 189695659,
            191912783, 387096133, 436273009, 1294268491, 1453168141,
            2300942549, 3842610773, 4302407359, 10726904659, 20678048297,
            22367084959, 25056082087, 42652618343, 127976334671, 182226896239,
            241160624143, 297501075799, 303371455241, 304599508537,
            416608695821, 461690510011, 614487453523, 738832927927,
            1346294310749, 1408695493609, 1968188556461, 2614941710599,
            7177162611713, 13829048559701, 19581334192423, 42842283925351,
            90874329411493, 171231342420521, 218209405436543, 1189459969825483,
            1686994940955803, 1693182318746371, 43841547845541059,
            55350776431903243, 80873624627234849, 203986478517455989,
            218034721194214273, 305405826521087869, 352521223451364323,
            401429925999153707, 418032645936712127, 804212830686677669,
            1425172824437699411, 5733241593241196731, 6787988999657777797
        }, /* clang-format on */
        {1,    2,    4,    6,    8,    14,   18,   20,   22,   34,   36,
         44,   52,   72,   86,   96,   112,  114,  118,  132,  148,  154,
         180,  210,  220,  222,  234,  248,  250,  282,  288,  292,  320,
         336,  354,  382,  384,  394,  456,  464,  468,  474,  486,  490,
         500,  514,  516,  532,  534,  540,  582,  588,  602,  652,  674,
         716,  766,  778,  804,  806,  906,  916,  924,  1132, 1184, 1198,
         1220, 1224, 1248, 1272, 1328, 1356, 1370, 1442, 1476, 1488, 1510}};
 
    return value;
}
 
// Returns pair (first_composite_in_gap, last_composite_in_gap).
// O(log(right)) approximately.
inline std::pair<uint64_t, uint64_t> prime_gap_upto(uint64_t right) {
    if (right < 4)
        throw detail::there_is_no_upto_error("prime gap", right);
 
    const auto &[P, G] = prime_gaps();
    for (int i = P.size() - 1;; --i) {
        if (P[i] >= right)
            continue;
 
        uint64_t real_right = std::min(right, P[i] + G[i] - 1);
        uint64_t prev = i > 0 ? G[i - 1] : 0;
        uint64_t curr = real_right - P[i];
 
        if (curr >= prev)
            return {P[i] + 1, real_right};
    }
}
 
// From https://oeis.org/A002182/b002182.txt.
inline const std::vector<uint64_t> &highly_composites() {
    /* clang-format off */
    static const std::vector<uint64_t> highly_composites = {
    1, 2, 4, 6, 12, 24, 36, 48, 60, 120, 180, 240, 360, 720, 840, 1260, 1680,
    2520, 5040, 7560, 10080, 15120, 20160, 25200, 27720, 45360, 50400, 55440,
    83160, 110880, 166320, 221760, 277200, 332640, 498960, 554400, 665280,
    720720, 1081080, 1441440, 2162160, 2882880, 3603600, 4324320, 6486480,
    7207200, 8648640, 10810800, 14414400, 17297280, 21621600, 32432400,
    36756720, 43243200, 61261200, 73513440, 110270160, 122522400, 147026880,
    183783600, 245044800, 294053760, 367567200, 551350800, 698377680, 735134400,
    1102701600, 1396755360, 2095133040, 2205403200, 2327925600, 2793510720,
    3491888400, 4655851200, 5587021440, 6983776800, 10475665200, 13967553600,
    20951330400, 27935107200, 41902660800, 48886437600, 64250746560,
    73329656400, 80313433200, 97772875200, 128501493120, 146659312800,
    160626866400, 240940299600, 293318625600, 321253732800, 481880599200,
    642507465600, 963761198400, 1124388064800, 1606268664000, 1686582097200,
    1927522396800, 2248776129600, 3212537328000, 3373164194400, 4497552259200,
    6746328388800, 8995104518400, 9316358251200, 13492656777600, 18632716502400,
    26985313555200, 27949074753600, 32607253879200, 46581791256000,
    48910880818800, 55898149507200, 65214507758400, 93163582512000,
    97821761637600, 130429015516800, 195643523275200, 260858031033600,
    288807105787200, 391287046550400, 577614211574400, 782574093100800,
    866421317361600, 1010824870255200, 1444035528936000, 1516237305382800,
    1732842634723200, 2021649740510400, 2888071057872000, 3032474610765600,
    4043299481020800, 6064949221531200, 8086598962041600, 10108248702552000,
    12129898443062400, 18194847664593600, 20216497405104000, 24259796886124800,
    30324746107656000, 36389695329187200, 48519593772249600, 60649492215312000,
    72779390658374400, 74801040398884800, 106858629141264000,
    112201560598327200, 149602080797769600, 224403121196654400,
    299204161595539200, 374005201994424000, 448806242393308800,
    673209363589963200, 748010403988848000, 897612484786617600,
    1122015605983272000, 1346418727179926400, 1795224969573235200,
    2244031211966544000, 2692837454359852800, 3066842656354276800,
    4381203794791824000, 4488062423933088000, 6133685312708553600,
    8976124847866176000, 9200527969062830400, 12267370625417107200ULL,
    15334213281771384000ULL, 18401055938125660800ULL}; /* clang-format on */
    return highly_composites;
}
 
// O(log(right)) approximately.
inline uint64_t highly_composite_upto(uint64_t right) {
    for (int i = highly_composites().size() - 1; i >= 0; --i)
        if (highly_composites()[i] <= right)
            return highly_composites()[i];
 
    throw detail::there_is_no_upto_error("highly composite number", right);
}
 
// O(log^3 (right)) expected.
// Generates a random prime in [left, right].
inline uint64_t gen_prime(uint64_t left, uint64_t right) {
    if (right < left or right < 2)
        throw detail::there_is_no_in_range_error("prime", left, right);
    left = std::max<uint64_t>(left, 2);
    auto [l_gap, r_gap] = prime_gap_upto(right);
    if (right - left + 1 <= r_gap - l_gap + 1) {
        // There might be no primes in the range.
        std::vector<uint64_t> vals(right - left + 1);
        iota(vals.begin(), vals.end(), left);
        shuffle(vals.begin(), vals.end());
        for (uint64_t i : vals)
            if (is_prime(i))
                return i;
        throw detail::there_is_no_in_range_error("prime", left, right);
    }
 
    uint64_t n;
    do {
        n = next(left, right);
    } while (!is_prime(n));
    return n;
}
 
// O(log^3 (left)) expected.
// left <= 2^64 - 59.
inline uint64_t prime_from(uint64_t left) {
    tgen_ensure(left <= std::numeric_limits<uint64_t>::max() - 58,
                "math: invalid bound");
    for (uint64_t i = std::max<uint64_t>(2, left);; ++i)
        if (is_prime(i))
            return i;
}
 
// O(log^3 (right)) expected.
inline uint64_t prime_upto(uint64_t right) {
    if (right >= 2)
        for (uint64_t i = right; i >= 2; --i)
            if (is_prime(i))
                return i;
    throw detail::there_is_no_upto_error("prime", right);
}
 
// O(n^(1/4) log n) expected.
// 0 < n.
inline int num_divisors(uint64_t n) {
    int divisors = 1;
    for (auto [p, e] : factor_by_prime(n))
        divisors *= (e + 1);
    return divisors;
}
 
// Random number in [left, right] with `divisor_count` divisors.
// O(log(right) log(divisor_count)).
// divisor_count must be prime.
inline uint64_t gen_divisor_count(uint64_t left, uint64_t right,
                                  int divisor_count) {
    tgen_ensure(divisor_count > 0 and is_prime(divisor_count),
                "math: divisor count must be prime");
    int root = divisor_count - 1;
    uint64_t lo = detail::kth_root_floor(left, root);
    if (*detail::expo(lo, root, left) < left)
        ++lo;
    uint64_t p = gen_prime(lo, detail::kth_root_floor(right, root));
    return *detail::expo(p, root, right);
}
 
// O(|mods| + log (right)).
// |rems| = |mods|.
// rems_i < mods_i.
inline uint64_t gen_congruent(uint64_t left, uint64_t right,
                              std::vector<uint64_t> rems,
                              std::vector<uint64_t> mods) {
    if (left > right)
        throw detail::there_is_no_in_range_error("congruent number", left,
                                                 right);
    tgen_ensure(rems.size() == mods.size(),
                "math: number of remainders and mods must be the same");
    tgen_ensure(rems.size() > 0, "math: must have at least one congruence");
 
    detail::crt crt;
    for (int i = 0; i < static_cast<int>(rems.size()); ++i) {
        tgen_ensure(rems[i] < mods[i],
                    "math: remainder must be smaller than the mod");
        crt = crt * detail::crt(rems[i], mods[i]);
 
        if (crt.a == -1)
            throw detail::there_is_no_in_range_error("congruent number", left,
                                                     right);
        if (crt.m > right) {
            if (!(left <= crt.a and crt.a <= right))
                throw detail::there_is_no_in_range_error("congruent number",
                                                         left, right);
 
            for (int j = 0; j < static_cast<int>(rems.size()); ++j)
                if (crt.a % mods[j] != rems[j])
                    throw detail::there_is_no_in_range_error("congruent number",
                                                             left, right);
            return crt.a;
        }
    }
 
    uint64_t k_min = crt.a >= left ? 0 : ((left - crt.a) + crt.m - 1) / crt.m;
    uint64_t k_max = (right - crt.a) / crt.m;
 
    if (k_min > k_max)
        throw detail::there_is_no_in_range_error("congruent number", left,
                                                 right);
 
    return crt.a + next(k_min, k_max) * crt.m;
}
 
// O(log (right)).
// rem < mod.
inline uint64_t gen_congruent(uint64_t left, uint64_t right, uint64_t rem,
                              uint64_t mod) {
    return gen_congruent(left, right, std::vector<uint64_t>({rem}),
                         std::vector<uint64_t>({mod}));
}
 
// First congruent number >= left.
// O(|mods| + log (left)).
// |rems| = |mods|.
// rems_i < mods_i.
inline uint64_t congruent_from(uint64_t left, std::vector<uint64_t> rems,
                               std::vector<uint64_t> mods) {
    tgen_ensure(rems.size() == mods.size(),
                "math: number of remainders and mods must be the same");
    tgen_ensure(rems.size() > 0, "math: must have at least one congruence");
 
    detail::crt crt;
    for (int i = 0; i < static_cast<int>(rems.size()); ++i) {
        tgen_ensure(rems[i] < mods[i],
                    "math: remainder must be smaller than the mod");
        crt = crt * detail::crt(rems[i], mods[i]);
 
        if (crt.a == -1)
            throw detail::there_is_no_from_error("congruent number", left);
        if (crt.m > std::numeric_limits<uint64_t>::max()) {
            if (crt.a < left)
                throw detail::error(
                    "math: congruent number does not exist or is too large");
 
            for (int j = 0; j < static_cast<int>(rems.size()); ++j)
                if (crt.a % mods[j] != rems[j])
                    throw detail::error("math: congruent number does "
                                        "not exist or is too large");
            return crt.a;
        }
    }
 
    uint64_t k = 0;
    if (crt.a < left)
        k = ((left - crt.a) + crt.m - 1) / crt.m;
    detail::i128 result = crt.a + k * crt.m;
 
    if (result > std::numeric_limits<uint64_t>::max())
        throw detail::error("math: congruent number is too large");
    return result;
}
 
// O(log (left))
// rem < mod.
inline uint64_t congruent_from(uint64_t left, uint64_t rem, uint64_t mod) {
    return congruent_from(left, std::vector<uint64_t>{rem},
                          std::vector<uint64_t>{mod});
}
 
// Last congruent number <= right.
// O(|mods| + log (right)).
// |rems| = |mods|.
// rems_i < mods_i.
inline uint64_t congruent_upto(uint64_t right, std::vector<uint64_t> rems,
                               std::vector<uint64_t> mods) {
    tgen_ensure(rems.size() == mods.size(),
                "math: number of remainders and mods must be the same");
    tgen_ensure(rems.size() > 0, "math: must have at least one congruence");
 
    detail::crt crt;
    for (int i = 0; i < static_cast<int>(rems.size()); ++i) {
        tgen_ensure(rems[i] < mods[i],
                    "math: remainder must be smaller than the mod");
 
        crt = crt * detail::crt(rems[i], mods[i]);
 
        if (crt.a == -1)
            throw detail::there_is_no_upto_error("congruent number", right);
        if (crt.m > right) {
            if (!(crt.a <= right))
                throw detail::there_is_no_upto_error("congruent number", right);
 
            for (int j = 0; j < static_cast<int>(rems.size()); ++j)
                if (crt.a % mods[j] != rems[j])
                    throw detail::there_is_no_upto_error("congruent number",
                                                         right);
            return crt.a;
        }
    }
 
    if (crt.a > right)
        throw detail::there_is_no_upto_error("congruent number", right);
 
    uint64_t k = (right - crt.a) / crt.m;
    detail::i128 result = crt.a + k * crt.m;
 
    if (result < 0)
        throw detail::there_is_no_upto_error("congruent number", right);
    return result;
}
 
// O(log r)
// rem < mod.
inline uint64_t congruent_upto(uint64_t right, uint64_t rem, uint64_t mod) {
    return congruent_upto(right, std::vector<uint64_t>{rem},
                          std::vector<uint64_t>{mod});
}
 
// Mod used for FFT/NTT.
inline constexpr int FFT_MOD = 998244353;
 
// Fibonacci sequence up to 2^64.
inline const std::vector<uint64_t> &fibonacci() {
    static const std::vector<uint64_t> fib = [] {
        std::vector<uint64_t> v = {0, 1};
        while (v.back() <=
               std::numeric_limits<uint64_t>::max() - v[v.size() - 2])
            v.push_back(v.back() + v[v.size() - 2]);
        return v;
    }();
    return fib;
}
 
// Partition is ordered (composition), that is, (1, 1, 2) != (1, 2, 1).
// O(n).
// 0 < n.
// 0 < part_left.
inline std::vector<int>
gen_partition(int n, int part_left = 1,
              std::optional<int> part_right = std::nullopt) {
    if (!part_right.has_value())
        part_right = n;
    part_right = std::min(*part_right, n);
    tgen_ensure(n > 0 and part_left > 0,
                "math: invalid parameters to gen_partition");
    tgen_ensure(part_left <= n and *part_right > 0, "math: no such partition");
 
    // dp[i] = log(number of ways to add to i).
    std::vector<long double> dp(n + 1, detail::LOG_ZERO);
    dp[0] = detail::LOG_ONE;
    long double window = detail::LOG_ZERO;
    for (int i = 1; i <= n; ++i) {
        if (i >= part_left)
            window = detail::add_log_space(window, dp[i - part_left]);
        if (i >= *part_right + 1)
            window = detail::sub_log_space(window, dp[i - *part_right - 1]);
        dp[i] = window;
    }
    tgen_ensure(dp[n] >= 0, "math: no such partition");
 
    // Crazy math tricks ahead.
    auto dp_pref = dp;
    for (int i = 1; i <= n; ++i)
        dp_pref[i] = detail::add_log_space(dp_pref[i - 1], dp[i]);
 
    std::vector<int> part;
    int sum = n;
    while (sum > 0) {
        // Will generate a number such that what remains is in [l, r].
        int l = std::max(0, sum - *part_right), r = sum - part_left;
        detail::tgen_ensure_against_bug(r >= 0, "math: r < 0 in gen_partition");
 
        int nxt_sum = std::min(sum, r);
        long double random = next<long double>(0, 1);
 
        // We generate a value X (log space), and then choose nxt_sum such
        // that dp_pref[nxt_sum-1] < X <= dp_pref[nxt_sum].
 
        // Math hack:
        // Let A = pref[l-1], B = pref[r], U = rand().
        // X = log[exp(A) + U * (exp(B) - exp(A))]
        //   = log{exp(B) * [exp(A) / exp(B) + U * (1 - exp(A) / exp(B))]}
        //   = B + log[exp(A - B) + U - U * exp(A - B))]
        //   = B + log[U + (1 - U) * exp(A - B)].
        long double val_l = l ? dp_pref[l - 1] : detail::LOG_ZERO,
                    val_r = dp_pref[r];
        while (nxt_sum > l and
               dp_pref[nxt_sum - 1] >=
                   val_r + detail::log_space(random +
                                             (1 - random) * exp(val_l - val_r)))
            --nxt_sum;
 
        part.push_back(sum - nxt_sum);
        sum = nxt_sum;
    }
 
    return part;
}
 
// Partition is ordered (composition), that is, (1, 1, 2) != (1, 2, 1).
// O(n) time/memory if part_right is not set, O(n * k) time/memory otherwise.
// 0 < k <= n.
// 0 <= part_left.
inline std::vector<int>
gen_partition_fixed_size(int n, int k, int part_left = 0,
                         std::optional<int> part_right = std::nullopt) {
    if (!part_right.has_value())
        part_right = n;
    part_right = std::min(*part_right, n);
    tgen_ensure(0 < k and k <= n and part_left >= 0,
                "math: invalid parameters to gen_partition_fixed_size");
    tgen_ensure(static_cast<long long>(k) * part_left <= n and
                    n <= static_cast<long long>(k) * (*part_right),
                "math: no such partition");
 
    // What we need to distribute to the parts.
    int s = n - k * part_left;
 
    std::vector<int> part(k);
    if (*part_right == n) {
        // Stars and bars - O(n).
        std::vector<int> cuts = {-1};
 
        int total = s + k - 1, bars = k - 1;
        for (int i = 0; i < total and bars > 0; ++i)
            if (next<long double>(0, 1) <
                static_cast<long double>(bars) / (total - i)) {
                cuts.push_back(i);
                --bars;
            }
        cuts.push_back(total);
 
        // Recovers parts.
        for (int i = 0; i < k; ++i)
            part[i] = cuts[i + 1] - cuts[i] - 1;
    } else {
        // DP with log trick - O(nk).
        int u = *part_right - part_left;
 
        // dp[i][j] = log(#ways to fill i parts with sum j)
        std::vector<std::vector<long double>> dp(
            k + 1, std::vector<long double>(s + 1, detail::LOG_ZERO));
        dp[0][0] = detail::LOG_ONE;
 
        for (int i = 1; i <= k; ++i) {
            std::vector<long double> pref = dp[i - 1];
            for (int j = 1; j <= s; ++j)
                pref[j] = detail::add_log_space(pref[j - 1], dp[i - 1][j]);
 
            for (int j = 0; j <= s; ++j) {
                dp[i][j] = pref[j];
                if (j >= u + 1)
                    dp[i][j] = detail::sub_log_space(dp[i][j], pref[j - u - 1]);
            }
        }
 
        // Recovers parts backwards.
        int left_to_distribute = s;
        for (int i = k; i >= 1; --i) {
            long double log_total = detail::LOG_ZERO;
            for (int j = 0; j <= u and j <= left_to_distribute; ++j)
                log_total = detail::add_log_space(
                    log_total, dp[i - 1][left_to_distribute - j]);
            detail::tgen_ensure_against_bug(
                log_total != detail::LOG_ZERO,
                "math: total == 0 in gen_partition_fixed_size");
 
            // Now we choose a number with probability proportional to
            // dp[i-1][.].
 
            // log(rand() * total) = log(rand()) + log(total).
            long double random =
                detail::log_space(next<long double>(0, 1)) + log_total;
 
            long double cur_prob = detail::LOG_ZERO;
            int chosen = 0;
            for (int j = 0; j <= u and j <= left_to_distribute; ++j) {
                cur_prob = detail::add_log_space(
                    cur_prob, dp[i - 1][left_to_distribute - j]);
                if (random < cur_prob) {
                    chosen = j;
                    break;
                }
            }
 
            part[k - i] = chosen;
            left_to_distribute -= chosen;
        }
    }
 
    for (int &i : part)
        i += part_left;
    return part;
}
 
// Partition is ordered (composition), that is, (1, 1, 2) != (1, 2, 1).
// Inspired by jngen rndm.partition: random delimiters, sort, gap recovery;
// omits jngen's part reordering, shuffles, and two-pass redistribution.
// 0 < k <= n.
// 0 <= part_left.
// Not uniformly random; optimized for speed.
// O(k log k).
inline std::vector<uint64_t> gen_partition_fixed_size_fast(
    uint64_t n, int k, uint64_t part_left = 0,
    std::optional<uint64_t> part_right = std::nullopt) {
    if (!part_right.has_value())
        part_right = n;
    part_right = std::min(*part_right, n);
 
    detail::u128 n128 = n;
    detail::u128 k128 = k;
    detail::u128 part_left128 = part_left;
    detail::u128 part_right128 = *part_right;
 
    tgen_ensure(k > 0 and k128 <= n128,
                "math: invalid parameters to gen_partition_fixed_size_fast");
    tgen_ensure(part_right128 >= part_left128 and
                    k128 * part_left128 <= n128 and
                    k128 * part_right128 >= n128,
                "math: no such partition");
 
    uint64_t slack_total = n128 - k128 * part_left128;
    uint64_t slack_max = part_right128 - part_left128;
 
    std::vector<uint64_t> part(k);
    if (k == 1) {
        part[0] = slack_total;
    } else {
        std::vector<uint64_t> cuts(k - 1);
        for (uint64_t &d : cuts)
            d = next<uint64_t>(0, slack_total);
        std::sort(cuts.begin(), cuts.end());
 
        uint64_t prev = 0;
        for (int i = 0; i + 1 < k; ++i) {
            part[i] = cuts[i] - prev;
            prev = cuts[i];
        }
        part[k - 1] = slack_total - prev;
    }
 
    auto add_part_left = [part_left](uint64_t x) -> uint64_t {
        detail::u128 val = x + part_left;
        detail::tgen_ensure_against_bug(
            val <= std::numeric_limits<uint64_t>::max(),
            "math: part + part_left exceeds uint64_t in "
            "gen_partition_fixed_size_fast");
        return val;
    };
 
    if (slack_max >= slack_total) {
        for (uint64_t &x : part)
            x = add_part_left(x);
        return part;
    }
 
    detail::u128 remaining = 0;
    for (uint64_t &x : part) {
        if (x > slack_max) {
            remaining += x - slack_max;
            x = slack_max;
        }
        x = add_part_left(x);
    }
 
    if (remaining > 0) {
        for (uint64_t &x : part) {
            if (x < *part_right && remaining > 0) {
                detail::u128 room = *part_right - x;
                detail::u128 add = std::min(remaining, room);
                detail::u128 val = x + add;
                detail::tgen_ensure_against_bug(
                    val <= *part_right,
                    "math: part exceeds part_right after redistribution in "
                    "gen_partition_fixed_size_fast");
                x = val;
                remaining -= add;
            }
        }
        detail::tgen_ensure_against_bug(
            remaining == 0, "math: remaining mass after redistribution in "
                            "gen_partition_fixed_size_fast");
    }
 
    return part;
}
 
// Random partition of elements into k ordered groups (input order preserved).
// If max_size is unset, part sizes are uniform via gen_partition_fixed_size.
// If max_size is set, uses gen_partition_fixed_size_fast (not uniform).
// O(n) if max_size is unset; O(n + k log k) if max_size is set.
template <typename T>
std::vector<std::vector<T>>
partition_elements(std::vector<T> elements, int k, int min_size = 0,
                   std::optional<uint64_t> max_size = std::nullopt) {
    size_t n = elements.size();
    tgen_ensure(k > 0, "math: partition_elements: k must be positive");
    tgen_ensure(min_size >= 0,
                "math: partition_elements: min_size must be non-negative");
 
    std::vector<uint64_t> sizes;
    if (max_size.has_value()) {
        sizes = gen_partition_fixed_size_fast(n, k, min_size, max_size);
    } else {
        for (int sz : gen_partition_fixed_size(n, k, min_size))
            sizes.push_back(sz);
    }
 
    std::vector<std::vector<T>> groups;
    groups.reserve(k);
    size_t pos = 0;
    for (uint64_t sz : sizes) {
        groups.emplace_back(elements.begin() + pos,
                            elements.begin() + pos + sz);
        pos += sz;
    }
    return groups;
}
 
}; // namespace math
 
/**************
 *            *
 *   STRING   *
 *            *
 **************/
 
namespace detail {
 
/*
 * Regex.
 *
 * Compatible with testlib's regex.
 *
 * Operations:
 * - A single character yields itself ("a", "3").
 * - A list of characters inside square braces yields any a random element
 *   from the list ("[abc123]").
 * - A range of characters is equivalent to listing them ("[a-z1-9A-Z]").
 * - A pattern followed by {n} yields the pattern repeated n times ("a{3}").
 * - A pattern followed by {l,r} yields the pattern repeated between l and r
 *   times, uniformly at random ("a{3,5}").
 * - A list of patterns separated by | yields a random pattern from the
 *   list, uniformly at random ("abc|def|ghi").
 * - Parentheses can be used for grouping ("a((a|b){3})").
 *
 * Examples:
 * 1. str("[1-9][0-9]{1,2}") generates two- or three-digit numbers.
 * 2. str("a[b-d]{2}|e") generates "e" or a random string of length 3, with
 *                       the first character being 'a' and the second and
 *                       third characters being 'b', 'c', or 'd'.
 * 3. str("[1-9][0-9]{%d}", n-1) generates n-digit numbers.
 *
 * Operations defined by {n} and {l,r} are applied from left to right, and
 * the pattern that comes before has its delimiters defined either by () or
 * [] at its end or is taken from the beginning of the pattern (in
 * "a[bc]{2}", "{2}" is applied to "[bc]", and in "[01]abc{3}", the "{3}" is
 * applied to "[01]abc").
 */
 
// If it has children, it is either a SEQ or an OR group, defined by the
// pattern_ field.
struct regex_node {
    // Considered to be repetition of left_bound != -1, pattern if
    // children_.empty(), otherwise "SEQ" or "OR", defined by the pattern_
    // field.
    std::string
        pattern_; // Either pattern, or "SEQ" or "OR" (if !children_.empty()).
    std::vector<regex_node> children_; // Children, when SEQ or OR.
    int left_bound_, right_bound_; // Left and right bounds of the repetition,
                                   // or -1 if not a repetition.
    double
        log_space_num_ways_; // Log space number of ways to match the pattern.
    std::optional<distinct_container<char>>
        distinct_; // Distinct generator for the pattern, for [chars].
 
    // c or [chars].
    regex_node(const std::string &pattern)
        : pattern_(pattern), left_bound_(-1), right_bound_(-1) {
        if (pattern.size() == 1) {
            log_space_num_ways_ = math::detail::LOG_ONE;
            return;
        }
        tgen_ensure_against_bug(pattern[0] == '[' and pattern.back() == ']',
                                "str: invalid regex: expected character class");
        int size = pattern.size() - 2;
        log_space_num_ways_ = math::detail::log_space(size);
        distinct_ = distinct_container<char>(pattern.substr(1, size));
    }
    // SEQ or OR.
    regex_node(const std::string &pattern, std::vector<regex_node> &children)
        : pattern_(pattern), left_bound_(-1), right_bound_(-1) {
        if (pattern == "SEQ") {
            // Multiply the number of ways.
            log_space_num_ways_ = math::detail::LOG_ONE;
            for (const auto &child : children)
                log_space_num_ways_ += child.log_space_num_ways_;
        } else if (pattern == "OR") {
            // Add the number of ways.
            log_space_num_ways_ = math::detail::LOG_ZERO;
            for (const auto &child : children)
                log_space_num_ways_ = math::detail::add_log_space(
                    log_space_num_ways_, child.log_space_num_ways_);
        } else
            tgen_ensure_against_bug("str: invalid regex: expected SEQ or OR");
 
        children_ = std::move(children);
        children.clear();
    }
    // REP.
    regex_node(int left_bound, int right_bound, regex_node &child)
        : pattern_("REP"), left_bound_(left_bound), right_bound_(right_bound) {
        log_space_num_ways_ = math::detail::LOG_ZERO;
        for (int i = left_bound; i <= right_bound; ++i)
            log_space_num_ways_ = math::detail::add_log_space(
                log_space_num_ways_, i * child.log_space_num_ways_);
 
        children_.push_back(std::move(child));
    }
};
 
// State of the regex parser.
struct regex_state {
    std::vector<regex_node> cur;      // Current sequence of nodes.
    std::vector<regex_node> branches; // Branches of the current OR group.
};
 
// Creates a SEQ node from the current state.
inline regex_node make_regex_seq(regex_state &st) {
    return regex_node("SEQ", st.cur);
}
 
// Finishes current state.
inline regex_node finish_regex_state(regex_state &st) {
    // SEQ.
    if (st.branches.empty())
        return make_regex_seq(st);
 
    // OR.
    st.branches.push_back(make_regex_seq(st));
    return regex_node("OR", st.branches);
}
 
// Parses a regex pattern into a tree, computing the number of ways to match the
// pattern.
inline regex_node parse_regex(std::string regex) {
    std::string new_regex;
    for (char c : regex)
        if (c != ' ')
            new_regex += c;
    swap(regex, new_regex);
    regex_state cur;
    std::vector<regex_state> stack;
 
    for (size_t i = 0; i < regex.size(); ++i) {
        char c = regex[i];
 
        if (c == '(') {
            // Pushes the current state to the stack.
            stack.push_back(std::move(cur));
            cur = regex_state();
        } else if (c == ')') {
            // Finishes the current state, and adds it to the parent.
            regex_node node = finish_regex_state(cur);
 
            tgen_ensure(!stack.empty(), "str: invalid regex: unmatched `)`");
            cur = std::move(stack.back());
            stack.pop_back();
 
            cur.cur.push_back(std::move(node));
        } else if (c == '|') {
            // Starts a new OR group.
            regex_node node = make_regex_seq(cur);
            cur.branches.push_back(std::move(node));
        } else if (c == '[') {
            // Parses a character class.
            std::string chars;
 
            for (++i; i < regex.size() and regex[i] != ']'; ++i) {
                if (i + 2 < regex.size() and regex[i + 1] == '-') {
                    char a = regex[i], b = regex[i + 2];
                    if (a > b)
                        std::swap(a, b);
                    for (char x = a; x <= b; ++x)
                        chars += x;
                    i += 2;
                } else
                    chars += regex[i];
            }
 
            tgen_ensure(i < regex.size() and regex[i] == ']',
                        "str: invalid regex: unmatched `[`");
            cur.cur.emplace_back("[" + chars + "]");
        } else if (c == '{') {
            // Parses a repetition.
            ++i;
            int l = -1, r = -1;
 
            while (i < regex.size() and
                   isdigit(static_cast<unsigned char>(regex[i]))) {
                if (l == -1)
                    l = 0;
                tgen_ensure(l <= static_cast<int>(1e8),
                            "str: invalid regex: number too large inside `{}`");
                l = 10 * l + (regex[i] - '0');
                ++i;
            }
 
            if (i < regex.size() and regex[i] == ',') {
                ++i;
                while (i < regex.size() and
                       isdigit(static_cast<unsigned char>(regex[i]))) {
                    if (r == -1)
                        r = 0;
                    tgen_ensure(
                        r <= static_cast<int>(1e8),
                        "str: invalid regex: number too large inside `{}`");
                    r = 10 * r + (regex[i] - '0');
                    ++i;
                }
            } else
                r = l;
 
            tgen_ensure(i < regex.size() and regex[i] == '}',
                        "str: invalid regex: unmatched `{`");
            tgen_ensure(l != -1 and r != -1,
                        "str: invalid regex: missing number inside `{}`");
            tgen_ensure(l <= r,
                        "str: invalid regex: invalid range inside `{}`");
 
            // Creates a REP node from the previous node.
            tgen_ensure(!cur.cur.empty(),
                        "str: invalid regex: expected expression before `{}`");
 
            regex_node rep(l, r, cur.cur.back());
            cur.cur.pop_back();
            cur.cur.push_back(std::move(rep));
        } else {
            // Creates a char node.
            cur.cur.emplace_back(std::string(1, c));
        }
    }
 
    tgen_ensure(stack.empty(), "str: invalid regex: unmatched `(`");
    return finish_regex_state(cur);
}
 
// Generates a uniformly random string that matches the given regex.
inline void gen_regex(const regex_node &node, std::string &str) {
    // For [chars], generate a random character from the list.
    if (node.pattern_[0] == '[') {
        str += node.pattern_[1 + next<int>(0, node.pattern_.size() - 3)];
        return;
    }
 
    // For REP, generate a random number of times to repeat the pattern.
    if (node.left_bound_ != -1) {
        // Generates a random value W from 0 to num_ways.
        // log(W) = log(random(0, 1) * num_ways)
        //        = log(random(0, 1)) + log(num_ways).
        double log_rand = math::detail::log_space(next<double>(0, 1)) +
                          node.log_space_num_ways_;
        double cur_prob = math::detail::LOG_ZERO;
        double child_num_ways = node.children_[0].log_space_num_ways_;
 
        for (int i = node.left_bound_; i <= node.right_bound_; ++i) {
            cur_prob =
                math::detail::add_log_space(cur_prob, i * child_num_ways);
            if (log_rand <= cur_prob) {
                for (int j = 0; j < i; ++j)
                    gen_regex(node.children_[0], str);
                return;
            }
        }
 
        tgen_ensure_against_bug(false,
                                "str: log_rand > cur_prob in REP gen_regex");
    }
 
    // For SEQ, generate all children.
    if (!node.children_.empty() and node.pattern_ == "SEQ") {
        for (const regex_node &child : node.children_)
            gen_regex(child, str);
        return;
    }
 
    // For OR, generate a random child.
    if (!node.children_.empty() and node.pattern_ == "OR") {
        // Generates a random value W from 0 to num_ways.
        // log(W) = log(random(0, 1) * num_ways)
        //        = log(random(0, 1)) + log(num_ways).
        double log_rand = math::detail::log_space(next<double>(0, 1)) +
                          node.log_space_num_ways_;
        double cur_prob = math::detail::LOG_ZERO;
 
        for (const regex_node &child : node.children_) {
            cur_prob = math::detail::add_log_space(cur_prob,
                                                   child.log_space_num_ways_);
            if (log_rand <= cur_prob) {
                gen_regex(child, str);
                return;
            }
        }
 
        tgen_ensure_against_bug(false,
                                "str: log_rand > cur_prob in OR gen_regex");
    }
 
    // For char, generate the character.
    detail::tgen_ensure_against_bug(
        node.pattern_.size() == 1,
        "str: invalid regex: expected single character, but got `" +
            node.pattern_ + "`");
    str += node.pattern_[0];
}
 
// Formats a regex string with given arguments.
template <typename... Args>
std::string regex_format(const std::string &s, Args &&...args) {
    if constexpr (sizeof...(Args) == 0) {
        return s;
    } else {
        int size = std::snprintf(nullptr, 0, s.c_str(), args...) + 1;
        std::string buf(size, '\0');
        std::snprintf(buf.data(), size, s.c_str(), args...);
        buf.pop_back(); // remove '\0'
        return buf;
    }
}
 
} // namespace detail
 
/*
 * String generator.
 */
 
struct str : gen_base<str> {
    std::optional<list<char>> list_; // List of characters.
    std::optional<detail::regex_node>
        root_; // Root node of the regex tree for the whole string.
 
    // Creates generator for strings of size 'size', with random characters in
    // [value_left, value_right].
    str(int size, char value_left = 'a', char value_right = 'z') {
        tgen_ensure(size > 0, "str: size must be positive");
        list_ = list<char>(size, value_left, value_right);
    }
 
    // Creates generator for strings of size 'size', with random characters in
    // 'chars'.
    str(int size, std::set<char> chars) {
        tgen_ensure(size > 0, "str: size must be positive");
        list_ = list<char>(size, chars);
    }
 
    // Creates generator for strings that match the given regex.
    template <typename... Args> str(const std::string &regex, Args &&...args) {
        tgen_ensure(regex.size() > 0, "str: regex must be non-empty");
 
        root_ = detail::parse_regex(
            detail::regex_format(regex, std::forward<Args>(args)...));
    }
 
    // Restricts strings for str[idx] = value.
    str &fix(int idx, char character) {
        tgen_ensure(!root_, "str: cannot add restriction for regex");
        list_->fix(idx, character);
        return *this;
    }
 
    // Restricts strings for list[S] to be equal, for given subset S of indices.
    str &equal(std::set<int> indices) {
        tgen_ensure(!root_, "str: cannot add restriction for regex");
        list_->equal(indices);
        return *this;
    }
 
    // Restricts strings for str[idx_1] = str[idx_2].
    str &equal(int idx_1, int idx_2) {
        tgen_ensure(!root_, "str: cannot add restriction for regex");
        list_->equal(idx_1, idx_2);
        return *this;
    }
 
    // Restricts strings for str[left..right] to have all equal values.
    str &equal_range(int left, int right) {
        tgen_ensure(!root_, "str: cannot add restriction for regex");
        list_->equal_range(left, right);
        return *this;
    }
 
    // Restricts strings for all equal chars.
    str &all_equal() {
        tgen_ensure(!root_, "str: cannot add restriction for regex");
        list_->all_equal();
        return *this;
    }
 
    // Restricts strings for str[left..right] to be a palindrome.
    str &palindrome(int left, int right) {
        tgen_ensure(!root_, "str: cannot add restriction for regex");
        tgen_ensure(0 <= left and left <= right and right < list_->size_,
                    "str: range indices must be valid");
        for (int i = left; i < right - (i - left); ++i)
            equal(i, right - (i - left));
        return *this;
    }
 
    // Restricts strings for the entire string to be a palindrome.
    str &palindrome() {
        tgen_ensure(!root_, "str: cannot add restriction for regex");
        return palindrome(0, list_->size_ - 1);
    }
 
    // Restricts strings for str[S] to be different (distinct), for given subset
    // S of indices.
    str &different(std::set<int> indices) {
        tgen_ensure(!root_, "str: cannot add restriction for regex");
        list_->different(indices);
        return *this;
    }
 
    // Restricts strings for str[idx_1] != str[idx_2].
    str &different(int idx_1, int idx_2) {
        tgen_ensure(!root_, "str: cannot add restriction for regex");
        list_->different(idx_1, idx_2);
        return *this;
    }
 
    // Restricts lists for list[left..right] to have all different chars.
    str &different_range(int left, int right) {
        tgen_ensure(!root_, "str: cannot add restriction for regex");
        list_->different_range(left, right);
        return *this;
    }
 
    // Restricts strings for all chars to be different.
    str &all_different() {
        tgen_ensure(!root_, "str: cannot add restriction for regex");
        list_->all_different();
        return *this;
    }
 
    // str value.
    struct value : gen_value_base<value> {
        using tgen_is_sequential_tag = detail::is_sequential_tag;
 
        using value_type = char;
        using std_type = std::string;
        std::string str_;
 
        value(const std::string &str) : str_(str) {
            tgen_ensure(!str_.empty(), "str: value: cannot be empty");
        }
 
        // Fetches size.
        int size() const { return str_.size(); }
 
        // Fetches position idx.
        char &operator[](int idx) {
            tgen_ensure(0 <= idx and idx < size(),
                        "str: value: index out of bounds");
            return str_[idx];
        }
        const char &operator[](int idx) const {
            tgen_ensure(0 <= idx and idx < size(),
                        "str: value: index out of bounds");
            return str_[idx];
        }
 
        // Sorts characters in non-decreasing order.
        // O(n log n).
        value &sort() {
            std::sort(str_.begin(), str_.end());
            return *this;
        }
 
        // Reverses string.
        // O(n).
        value &reverse() {
            std::reverse(str_.begin(), str_.end());
            return *this;
        }
 
        // Lowercases all characters.
        // O(n).
        value &lowercase() {
            for (char &c : str_)
                c = std::tolower(c);
            return *this;
        }
 
        // Uppercases all characters.
        // O(n).
        value &uppercase() {
            for (char &c : str_)
                c = std::toupper(c);
            return *this;
        }
 
        // Concatenates two values.
        // Linear.
        value operator+(const value &rhs) const {
            return value(str_ + rhs.str_);
        }
 
        // Shuffles string uniformly.
        // O(n).
        value &shuffle() {
            for (int i = 0; i < size(); ++i)
                std::swap(str_[i], str_[next(0, size() - 1)]);
            return *this;
        }
 
        // Returns a random character uniformly.
        // O(1).
        char pick() const { return str_[next<int>(0, size() - 1)]; }
 
        // Returns str_[i] with probability proportional to distribution[i].
        // O(1).
        template <typename Dist>
        char pick_by_distribution(const std::vector<Dist> &distribution) const {
            tgen_ensure(static_cast<size_t>(size()) == distribution.size(),
                        "value and distribution must have the same size");
            return str_[next_by_distribution(distribution)];
        }
        template <typename Dist>
        char pick_by_distribution(
            const std::initializer_list<Dist> &distribution) const {
            return pick_by_distribution(std::vector<Dist>(distribution));
        }
 
        // Chooses k characters uniformly, as in a subsequence of size k.
        // O(n).
        value choose(int k) const {
            tgen_ensure(0 < k and k <= size(),
                        "number of elements to choose must be valid");
            std::string new_str;
            int need = k;
            for (int i = 0; need > 0; ++i) {
                int left = size() - i;
                if (next(1, left) <= need) {
                    new_str.push_back(str_[i]);
                    need--;
                }
            }
            return value(new_str);
        }
 
        // Prints to std::ostream.
        friend std::ostream &operator<<(std::ostream &out, const value &val) {
            return out << val.str_;
        }
 
        // Gets a std::string representing the value.
        std::string to_std() const { return std_type(str_); }
    };
 
    // Generates str value.
    // If created from restrictions: O(n log n).
    // If created from regex: expected linear.
    value gen() const {
        if (root_) {
            // Regex.
            std::string ret_str;
            gen_regex(*root_, ret_str);
            return value(ret_str);
        } else {
            // List.
            std::vector<char> vec = list_->gen().to_std();
            return value(std::string(vec.begin(), vec.end()));
        }
    }
};
 
/************
 *          *
 *   PAIR   *
 *          *
 ************/
 
namespace detail {
 
// Generates pair first == second.
// O(1).
template <typename T> std::pair<T, T> gen_eq(T L1, T R1, T L2, T R2) {
    T L = std::max(L1, L2);
    T R = std::min(R1, R2);
 
    tgen_ensure(L <= R, "pair: no valid values to generate");
    T x = next<T>(L, R);
    return {x, x};
}
 
// Returns {R1-L1+1, R2-L2+1}.
template <typename T>
std::pair<u128, u128> get_n_and_m(T L1, T R1, T L2, T R2) {
    u128 n = static_cast<i128>(R1) - L1 + 1;
    u128 m = static_cast<i128>(R2) - L2 + 1;
    return {n, m};
}
 
// Returns first + first+1 + ... + last,
// num_terms terms. Avoids overflow.
static u128 pos_arith_sum(u128 first, u128 last, u128 num_terms) {
    u128 x = first + last, y = num_terms;
 
    // x * y / 2, avoiding overflow.
    if (x % 2 == 0)
        x /= 2;
    else
        y /= 2;
 
    return x * y;
}
 
// Generates pair first != second.
// O(1) expected.
template <typename T> std::pair<T, T> gen_neq(T L1, T R1, T L2, T R2) {
    auto [n, m] = get_n_and_m(L1, R1, L2, R2);
 
    T L_intersect = std::max(L1, L2);
    T R_intersect = std::min(R1, R2);
    u128 inter = static_cast<i128>(R_intersect) - L_intersect + 1;
 
    u128 total = n * m - inter;
    tgen_ensure(total > 0, "pair: no valid values to generate");
 
    // Runs O(1) expected times in the worst case.
    T a, b;
    do {
        a = next<T>(L1, R1);
        b = next<T>(L2, R2);
    } while (a == b);
 
    return {a, b};
}
 
// For lt, splits 'second' into two regions:
// 1) second <= R1 -> number of 'first' is (second - L1)
// 2) second >  R1 -> number of 'first' is (R1 - L1 + 1)
// Returns {count_region1, count_region2}.
// O(1).
template <typename T>
std::pair<u128, u128> count_lt_regions(T L1, T R1, T L2, T R2) {
    auto [n, m] = get_n_and_m(L1, R1, L2, R2);
 
    // 'second' must be >= L1 + 1.
    i128 L_second = std::max<i128>(L2, static_cast<i128>(L1) + 1);
    i128 R_second = R2;
 
    // Split point for 'second'.
    i128 split = std::min<i128>(R_second, R1);
 
    // Region 1: b in [L_second, split].
    u128 len1 = std::max<i128>(0, split - L_second + 1);
 
    u128 count_region1 = 0;
    if (len1 > 0) {
        // For b in [L_second, split], there are (b - L1) ways.
        i128 first = L_second - L1;
        i128 last = split - L1;
 
        // Arithmetic series first + (first + 1) + ... + last, len1 terms.
        count_region1 = pos_arith_sum(first, last, len1);
    }
 
    // Region 2: b > R1.
    // For b in [R1+1, R_second], there are 'n' ways.
    i128 L_second_region2 = std::max(L_second, static_cast<i128>(R1) + 1);
 
    u128 len2 = std::max<i128>(0, R_second - L_second_region2 + 1);
    u128 count_region2 = len2 * n;
 
    return {count_region1, count_region2};
}
 
// Generates pair first < second.
// O(log(R1 - L1 + 1) + log(R2 - L2 + 1)).
template <typename T> std::pair<T, T> gen_lt(T L1, T R1, T L2, T R2) {
    auto [n, m] = get_n_and_m(L1, R1, L2, R2);
 
    // 'second' needs to be at least L1 + 1 to have a valid value for
    // 'first'.
    i128 L_second = std::max<i128>(L2, static_cast<i128>(L1) + 1);
    i128 R_second = R2;
 
    // Splits 'second' into two regions:
    // 1) b <= R1 -> number of 'first' is (b - L1);
    // 2) b >  R1 -> number of 'first' is (R1 - L1 + 1).
    i128 split = std::min<i128>(R_second, R1);
 
    auto [count_region1, count_region2] = count_lt_regions(L1, R1, L2, R2);
    u128 total = count_region1 + count_region2;
    tgen_ensure(total > 0, "pair: no valid values to generate");
 
    u128 k = detail::next128(total);
    if (k < count_region1) {
        // Region 1: invert arithmetic series.
 
        // For b in [L_second, split].
        u128 len1 = std::max<i128>(0, split - L_second + 1);
 
        // We consider b in [L_second, L_second + d].
        // Each b contributes (b - L1) = base + (b - L_second).
        // So we sum: base + (base+1) + ... + (base+d)
        // d in [0, len1).
 
        i128 base = L_second - L1;
        i128 lo = 0, hi = static_cast<i128>(len1) - 1;
 
        while (lo < hi) {
            i128 mid = lo + (hi - lo) / 2;
 
            if (pos_arith_sum(base, base + mid, mid + 1) <= k)
                lo = mid + 1;
            else
                hi = mid;
        }
        i128 d = lo;
 
        // Subtracts prefix sum with d-1 terms from k.
        if (d > 0)
            k -= pos_arith_sum(base, base + d - 1, d);
 
        return {L1 + static_cast<T>(k), L_second + d};
    } else {
        // Region 2: uniform block of size n.
        k -= count_region1;
 
        // For b in [R1+1, R_second], there are 'n' ways.
        i128 L_second_region2 = std::max(L_second, static_cast<i128>(R1) + 1);
 
        return {L1 + static_cast<T>(k % n),
                L_second_region2 + static_cast<T>(k / n)};
    }
}
 
// Generates pair first > second.
// O(log(R1 - L1 + 1) + log(R2 - L2 + 1)).
template <typename T> std::pair<T, T> gen_gt(T L1, T R1, T L2, T R2) {
    auto [first, second] = gen_lt(L2, R2, L1, R1);
    return {second, first};
}
 
// Generates pair first <= second.
// O(log(R1 - L1 + 1) + log(R2 - L2 + 1)).
template <typename T> std::pair<T, T> gen_leq(T L1, T R1, T L2, T R2) {
    // Counts how many pairs are there with first = second.
    i128 L_intersect = std::max(L1, L2);
    i128 R_intersect = std::min(R1, R2);
    u128 eq_count = std::max<i128>(0, R_intersect - L_intersect + 1);
 
    // Counts how many pairs are there with first < second.
    auto [lt_region1, lt_region2] = count_lt_regions(L1, R1, L2, R2);
    u128 lt_count = lt_region1 + lt_region2;
 
    u128 total = eq_count + lt_count;
    tgen_ensure(total > 0, "pair: no valid values to generate");
 
    if (detail::next128(total) < eq_count)
        return gen_eq(L1, R1, L2, R2);
    return gen_lt(L1, R1, L2, R2);
}
 
// Generates pair first >= second.
// O(log(R1 - L1 + 1) + log(R2 - L2 + 1)).
template <typename T> std::pair<T, T> gen_geq(T L1, T R1, T L2, T R2) {
    auto [first, second] = gen_leq(L2, R2, L1, R1);
    return {second, first};
}
 
}; // namespace detail
 
/*
 * Pair generator.
 *
 * Pairs of integral types.
 */
 
template <typename T> struct pair : gen_base<pair<T>> {
    std::pair<T, T> first_, second_; // Range of first and second values.
    // Type of restriction.
    enum class restriction_type { eq, neq, lt, gt, leq, geq, unspecified };
    restriction_type type_ = restriction_type::unspecified;
 
    // Creates a pair with random values in [first_l, first_r] and [second_l,
    // second_r].
    pair(T first_left, T first_right, T second_left, T second_right)
        : first_(first_left, first_right), second_(second_left, second_right) {
        tgen_ensure(first_left <= first_right,
                    "pair: first range must be valid");
        tgen_ensure(second_left <= second_right,
                    "pair: second range must be valid");
    }
 
    // Creates a pair with random values in [both_l, both_r].
    pair(T both_left, T both_right)
        : pair(both_left, both_right, both_left, both_right) {}
 
    // Restricts pair for first = second.
    pair &eq() {
        type_ = restriction_type::eq;
        return *this;
    }
 
    // Restricts pair for first != second.
    pair &neq() {
        type_ = restriction_type::neq;
        return *this;
    }
 
    // Restricts pair for first < second.
    pair &lt() {
        type_ = restriction_type::lt;
        return *this;
    }
 
    // Restricts pair for first > second.
    pair &gt() {
        type_ = restriction_type::gt;
        return *this;
    }
 
    // Restricts pair for first <= second.
    pair &leq() {
        type_ = restriction_type::leq;
        return *this;
    }
 
    // Restricts pair for first >= second.
    pair &geq() {
        type_ = restriction_type::geq;
        return *this;
    }
 
    // Pair value.
    struct value : gen_value_base<value> {
        using value_type = T;
        using std_type = std::pair<T, T>;
 
        std::pair<T, T> pair_;
        char sep_;
 
        value(const std::pair<T, T> &pair) : pair_(pair), sep_(' ') {}
        value(const T &first, const T &second)
            : pair_(first, second), sep_(' ') {}
 
        T first() const { return pair_.first; }
        T second() const { return pair_.second; }
 
        // Sets the separator for the pair, for printing.
        value &separator(char sep) {
            sep_ = sep;
            return *this;
        }
 
        // Prints to std::ostream, separated by sep_.
        friend std::ostream &operator<<(std::ostream &out, const value &val) {
            return out << val.pair_.first << val.sep_ << val.pair_.second;
        }
 
        // Gets a std::pair representing the value.
        auto to_std() const {
            if constexpr (!detail::is_generator_value<T>::value) {
                return pair_;
            } else {
                std::pair<typename T::std_type, typename T::std_type> pair(
                    pair_.first.to_std(), pair_.second.to_std());
                return pair;
            }
        }
    };
 
    // Generates a random pair.
    // O(log(R1 - L1 + 1) + log(R2 - L2 + 1)).
    value gen() const {
        T L1 = first_.first, R1 = first_.second;
        T L2 = second_.first, R2 = second_.second;
 
        switch (type_) {
        case restriction_type::unspecified:
            return {next<T>(L1, R1), next<T>(L2, R2)};
        case restriction_type::eq:
            return detail::gen_eq<T>(L1, R1, L2, R2);
        case restriction_type::neq:
            return detail::gen_neq<T>(L1, R1, L2, R2);
        case restriction_type::lt:
            return detail::gen_lt<T>(L1, R1, L2, R2);
        case restriction_type::gt:
            return detail::gen_gt<T>(L1, R1, L2, R2);
        case restriction_type::leq:
            return detail::gen_leq<T>(L1, R1, L2, R2);
        case restriction_type::geq:
            return detail::gen_geq<T>(L1, R1, L2, R2);
        }
        throw detail::error("pair: unknown restriction type");
    }
};
 
/************
 *          *
 *   TREE   *
 *          *
 ************/
 
namespace detail {
 
// Generates edges from Prufer sequence.
// O(n).
inline std::vector<std::pair<int, int>> edges_from_prufer(std::vector<int> p) {
    int n = p.size() + 2;
 
    // Degrees.
    std::vector<int> d(n, 1);
    for (int i : p)
        d[i]++;
 
    // Adds last vertex.
    p.push_back(n - 1);
 
    // Finds first vertex with degree 1.
    int idx, u;
    idx = u = find(d.begin(), d.end(), 1) - d.begin();
 
    // Generates edges.
    std::vector<std::pair<int, int>> edges;
    for (int v : p) {
        edges.emplace_back(u, v);
        if (--d[v] == 1 and v < idx)
            u = v;
        else
            idx = u = find(d.begin() + idx + 1, d.end(), 1) - d.begin();
    }
    return edges;
}
 
// Disjoint set union (union-find) for connectivity queries.
struct dsu {
    std::vector<int> parent_;
    std::vector<unsigned char> rank_;
 
    // Creates a dsu with `n` elements, indexed from 0 to n-1.
    // Initially every element is in its own set.
    // O(n).
    dsu(int n) : parent_(n), rank_(n, 0) {
        for (int i = 0; i < n; ++i)
            parent_[i] = i;
    }
 
    // Adds new elements to the dsu, each in their own new set.
    // O(k) amortized.
    void add_elements(int k) {
        for (int i = 0; i < k; ++i) {
            int new_id = parent_.size();
            parent_.push_back(new_id);
            rank_.push_back(0);
        }
    }
 
    // Finds representative of set containing i.
    // O(alpha(n)) amortized, O(log n) worst case.
    int find(int i) {
        return parent_[i] == i ? i : parent_[i] = find(parent_[i]);
    }
 
    // Merges components of `a` and `b`. Returns if the sets were united, and
    // false if a and b were in the same set.
    // O(alpha(n)) amortized, O(log n) worst case.
    bool unite(int a, int b) {
        a = find(a);
        b = find(b);
        if (a == b)
            return false;
        if (rank_[a] > rank_[b])
            std::swap(a, b);
        parent_[a] = b;
        if (rank_[a] == rank_[b])
            ++rank_[b];
        return true;
    }
};
 
} // namespace detail
 
// Forward declaration of wgraph.
template <typename VWeight, typename EWeight> struct wgraph;
 
/*
 * Tree generator.
 *
 * Unrooted trees with `n` vertices, indexed from 0 to n-1.
 * These are unrooted undirected labeled trees, that is, isomorphism is not
 * taken into account. VWeight is the type of vertex weights, and EWeight is
 * the type of edge weights. Generator does not generate weights. The weights
 * are to be set in the wtree::value.
 */
 
template <typename VWeight, typename EWeight>
struct wtree : gen_base<wtree<VWeight, EWeight>> {
    int n_;                               // Number of vertices.
    std::set<std::pair<int, int>> edges_; // Edges that were set.
 
    // Creates tree generator with `n` vertices.
    // O(1).
    wtree(int n) : n_(n) {
        tgen_ensure(n > 0, "wtree: number of vertices must be positive");
    }
 
    // Adds edge between u and v (this edge must be generated).
    // O(log n).
    wtree &add_edge(int u, int v) {
        tgen_ensure(0 <= std::min(u, v) and std::max(u, v) < n_,
                    "wtree: vertices must be indexed in [0, n)");
        tgen_ensure(u != v, "wtree: cannot add self loop to tree");
 
        if (u > v)
            std::swap(u, v);
        edges_.emplace(u, v);
        return *this;
    }
 
    // Tree value.
    //
    // Edges are stored in both directions in adjacency list, but only u < v in
    // edge list.
    struct value : gen_value_base<value> {
        using std_type = std::pair<int, std::vector<std::set<int>>>;
 
        int n_;                                  // Number of vertices.
        std::vector<std::set<int>> adj_;         // Adjacency list.
        std::vector<std::pair<int, int>> edges_; // Edge list.
        bool add_1_;   // If should add 1 for printing vertex ids.
        bool print_n_; // If should print n.
        std::optional<int> print_parents_; // If should print in parent style
                                           // (stores the root).
        std::optional<std::vector<VWeight>> vertex_weights_; // Vertex weights.
        std::optional<std::vector<EWeight>>
            edge_weights_; // Edge weights (in same order as edges_).
        detail::dsu dsu_;  // Connectivity of current edges (for cycle checks).
 
        // Creates value from adjacency list.
        // O(n).
        value(const std::vector<std::set<int>> &adj)
            : n_(static_cast<int>(adj.size())), adj_(adj), add_1_(false),
              print_n_(false), dsu_(n_) {
            for (int u = 0; u < n_; ++u)
                for (auto v : adj[u]) {
                    tgen_ensure(
                        0 <= v and v < n_,
                        "wtree: value: vertices must be indexed in [0, n)");
                    // Symmetric adjacency: count each undirected edge once.
                    if (u < v) {
                        edges_.emplace_back(u, v);
                        tgen_ensure(
                            dsu_.unite(u, v),
                            "wtree: value: initial graph must form a tree");
                    }
                }
        }
 
        // Creates value from `n` and edge list.
        // O(n).
        value(int n, const std::vector<std::pair<int, int>> &edges)
            : n_(n), adj_(n), add_1_(false), print_n_(false), dsu_(n) {
            edges_.reserve(edges.size());
            for (auto [u, v] : edges) {
                tgen_ensure(0 <= std::min(u, v) and std::max(u, v) < n,
                            "wtree: value: vertices must be indexed in [0, n)");
                tgen_ensure(dsu_.unite(u, v),
                            "wtree: value: initial graph must form a tree");
                if (u > v)
                    std::swap(u, v);
                edges_.emplace_back(u, v);
                adj_[u].insert(v);
                adj_[v].insert(u);
            }
        }
        value(int n, const std::set<std::pair<int, int>> &edges)
            : value(n, std::vector<std::pair<int, int>>(edges.begin(),
                                                        edges.end())) {}
        value(int n, const std::initializer_list<std::pair<int, int>> &edges)
            : value(n, std::vector<std::pair<int, int>>(edges)) {}
 
        // Creates tree from graph via Kruskal-like random spanning tree.
        // Implemented after wgraph definition.
        // O(n + m alpha(n)).
        value(const typename wgraph<VWeight, EWeight>::value &g);
 
        // Weight type conversion.
        // O(n).
        template <typename NewVWeight, typename NewEWeight>
        typename wtree<NewVWeight, NewEWeight>::value
        convert_weight_types() const {
            tgen_ensure(!vertex_weights_.has_value() and
                            !edge_weights_.has_value(),
                        "wtree: value: cannot convert weight type after "
                        "assigning weights");
 
            typename wtree<NewVWeight, NewEWeight>::value new_tree(adj_);
            new_tree.add_1_ = add_1_;
            new_tree.print_n_ = print_n_;
            new_tree.print_parents_ = print_parents_;
            return new_tree;
        }
 
        // Fetches number of vertices.
        int n() const { return n_; }
 
        // Fetches a const ref. to adjacency list.
        const std::vector<std::set<int>> &adj() const { return adj_; }
 
        // Fetches a const ref. to edge list.
        const std::vector<std::pair<int, int>> &edges() const { return edges_; }
 
        // Fetches a const ref. to vertex weights.
        const std::optional<std::vector<VWeight>> &vertex_weights() const {
            return vertex_weights_;
        }
 
        // Fetches a const ref. to edge weights.
        const std::optional<std::vector<EWeight>> &edge_weights() const {
            return edge_weights_;
        }
 
        // Sets vertex weights.
        // O(n).
        template <typename NewVWeight = VWeight>
        typename wtree<NewVWeight, EWeight>::value set_vertex_weights(
            const std::vector<NewVWeight> &vertex_weights) const {
            tgen_ensure(static_cast<int>(vertex_weights.size()) == n(),
                        "wtree: value: must give `n` vertex weights");
 
            auto new_tree = convert_weight_types<NewVWeight, EWeight>();
            new_tree.vertex_weights_ = vertex_weights;
            return new_tree;
        }
 
        // Sets edge weights.
        // O(n).
        template <typename NewEWeight = EWeight>
        typename wtree<VWeight, NewEWeight>::value
        set_edge_weights(const std::vector<NewEWeight> &edge_weights) const {
            tgen_ensure(
                edge_weights.size() == edges().size(),
                "wtree: value: must give `edges().size()` edge weights");
 
            auto new_tree = convert_weight_types<VWeight, NewEWeight>();
            new_tree.edge_weights_ = edge_weights;
            return new_tree;
        }
 
        // Enables edge-weighted mode before adding weighted edges
        // incrementally. The tree must have no edges yet. O(1).
        value &edge_weighted() {
            tgen_ensure(edges().size() == 0,
                        "wtree: value: edge_weighted requires a tree with no "
                        "edges");
            tgen_ensure(!edge_weights_.has_value(),
                        "wtree: value: tree is already edge-weighted");
 
            edge_weights_ = std::vector<EWeight>();
            return *this;
        }
 
        // Adds 1 to vertex ids, for printing.
        // O(1).
        value &add_1() {
            add_1_ = true;
            return *this;
        }
 
        // Prints `n` on a new line before printing the tree.
        // O(1).
        value &print_n() {
            print_n_ = true;
            return *this;
        }
 
        // Prints the tree in parent style.
        // If root = -1, the root is considered to be 0, and its parent is not
        // printed. Otherwise, prints the parent of the root as -1. If root = n,
        // randomizes the root. O(1).
        value &print_parents(int root = -1) {
            tgen_ensure(root == -1 or (0 <= root and root < n()) or root == n(),
                        "wtree: value: root must be -1, `n`, or in [0, n)");
            print_parents_ = root;
            return *this;
        }
 
        // Shuffles the tree's vertex labels (except those in `indices`,
        // which keep their current label) and edge order. The change is
        // applied eagerly to the underlying adjacency list, edge list,
        // vertex weights and edge weights.
        // O(n).
        value &shuffle_except(std::set<int> indices) {
            // Builds the relabeling: for each vertex `i`, `new_label[i]` is
            // its new id. Vertices in `indices` keep their label; the others
            // are permuted among themselves.
            std::vector<int> new_label(n());
            std::vector<int> shuffled;
            for (int i = 0; i < n(); ++i) {
                if (indices.count(i))
                    new_label[i] = i;
                else
                    shuffled.push_back(i);
            }
            std::vector<int> targets = shuffled;
            tgen::shuffle(targets.begin(), targets.end());
            for (size_t k = 0; k < shuffled.size(); ++k)
                new_label[shuffled[k]] = targets[k];
 
            // Rewrites adjacency list with new labels.
            std::vector<std::set<int>> new_adj(n());
            for (int u = 0; u < n(); ++u)
                for (int v : adj_[u])
                    new_adj[new_label[u]].insert(new_label[v]);
            adj_ = std::move(new_adj);
 
            // Rewrites edges with new labels (canonical undirected order).
            for (auto &[u, v] : edges_) {
                u = new_label[u];
                v = new_label[v];
                if (u > v)
                    std::swap(u, v);
            }
 
            // Permutes vertex weights to match the new labels.
            if (vertex_weights_.has_value()) {
                std::vector<VWeight> new_vw(n());
                for (int i = 0; i < n(); ++i)
                    new_vw[new_label[i]] = (*vertex_weights_)[i];
                vertex_weights_ = std::move(new_vw);
            }
 
            // Rebuilds the dsu so future `add_edge` calls see the new labels.
            dsu_ = detail::dsu(n());
            for (auto [u, v] : edges_)
                dsu_.unite(u, v);
 
            // Shuffles edge order, keeping edge weights aligned.
 
            std::vector<int> perm(edges_.size());
            std::iota(perm.begin(), perm.end(), 0);
            tgen::shuffle(perm.begin(), perm.end());
 
            std::vector<std::pair<int, int>> new_edges;
            std::optional<std::vector<EWeight>> new_ew;
            if (edge_weights_.has_value())
                new_ew = std::vector<EWeight>();
            for (int i : perm) {
                new_edges.push_back(edges_[i]);
                if (new_ew.has_value())
                    new_ew->push_back((*edge_weights_)[i]);
            }
            edges_ = new_edges;
            if (new_ew.has_value())
                edge_weights_ = new_ew;
 
            return *this;
        }
 
        // Shuffles the tree's vertices and edge order.
        // O(n).
        value &shuffle() { return shuffle_except({}); }
 
        // Adds edge (u, v).
        // O(log n) amortized.
        value &add_edge(int u, int v, std::optional<EWeight> w = std::nullopt) {
            tgen_ensure(0 <= std::min(u, v) and std::max(u, v) < n(),
                        "wtree: value: vertex ids must be valid");
 
            if (u > v)
                std::swap(u, v);
 
            if (adj_[u].count(v))
                return *this;
 
            adj_[u].insert(v);
            adj_[v].insert(u);
            edges_.emplace_back(u, v);
            tgen_ensure(dsu_.unite(u, v),
                        "wtree: value: added edge must not create a cycle");
 
            if (w.has_value()) {
                tgen_ensure(edge_weights().has_value(),
                            "wtree: value: cannot add weighted edge to "
                            "edge-unweighted tree");
 
                edge_weights_->push_back(*w);
            } else
                tgen_ensure(!edge_weights().has_value(),
                            "wtree: value: cannot add unweighted edge to "
                            "edge-weighted tree");
 
            return *this;
        }
 
        // Links tree with another `rhs`, adding the edge between u (in left
        // tree) and v (in right tree). Ids for added vertices are updated
        // accordingly.
        // O(rhs.n + rhs.m * log n) amortized.
        value &link(const value &rhs, int new_u, int new_v,
                    std::optional<EWeight> new_w = std::nullopt) {
            tgen_ensure(0 <= new_u and new_u < n() and 0 <= new_v and
                            new_v < rhs.n(),
                        "wtree: value: vertex ids must be valid");
 
            // Edges from right-hand side.
            int shift = n();
            add_vertices(rhs.n(), rhs.vertex_weights());
            for (int i = 0; i < static_cast<int>(rhs.edges().size()); ++i) {
                auto [u, v] = rhs.edges()[i];
                add_edge(shift + u, shift + v,
                         rhs.edge_weights().has_value()
                             ? std::optional<EWeight>((*rhs.edge_weights())[i])
                             : std::nullopt);
            }
 
            // New edge.
            add_edge(new_u, shift + new_v, new_w);
 
            return *this;
        }
 
        // Glues the tree with another `rhs` such that index_pairs[i].first is
        // considered to be the same as index_pairs[i].second. Ids for added
        // vertices are updated accordingly.
        // O(rhs.n + rhs.m * log n) amortized.
        value &glue(const value &rhs,
                    std::set<std::pair<int, int>> index_pairs) {
            // Checks validity of indices.
            std::set<int> idx_left, idx_right;
            std::vector<int> right_id_to_left(rhs.n(), -1);
            for (auto [l, r] : index_pairs) {
                tgen_ensure(
                    0 <= l and l < n() and 0 <= r and r < rhs.n(),
                    "wtree: value: vertex indices to glue must be valid");
                tgen_ensure(idx_left.count(l) == 0 and idx_right.count(r) == 0,
                            "wtree: value: must not have repeated indices "
                            "on the same side to glue");
 
                idx_left.insert(l);
                idx_right.insert(r);
                right_id_to_left[r] = l;
            }
 
            // Computes new ids of right vertices.
            std::vector<int> new_right_id(rhs.n(), -1);
            int intersection_lt = 0;
            std::optional<std::vector<VWeight>> rhs_vertex_weights;
            for (int i = 0; i < rhs.n(); ++i) {
                if (right_id_to_left[i] != -1) {
                    // Is in intersection.
                    ++intersection_lt;
                    new_right_id[i] = right_id_to_left[i];
                } else {
                    // New id.
                    new_right_id[i] = n() + i - intersection_lt;
                    if (rhs.vertex_weights().has_value()) {
                        if (!rhs_vertex_weights.has_value())
                            rhs_vertex_weights = std::vector<VWeight>();
                        rhs_vertex_weights->push_back(
                            (*rhs.vertex_weights())[i]);
                    }
                }
            }
 
            // Adds new vertices and edges.
            add_vertices(rhs.n() - intersection_lt, rhs_vertex_weights);
            for (int i = 0; i < static_cast<int>(rhs.edges().size()); ++i) {
                auto [u, v] = rhs.edges()[i];
                add_edge(new_right_id[u], new_right_id[v],
                         rhs.edge_weights().has_value()
                             ? std::optional<EWeight>((*rhs.edge_weights())[i])
                             : std::nullopt);
            }
 
            return *this;
        }
        value &glue(const value &rhs,
                    std::initializer_list<std::pair<int, int>> il) {
            return glue(rhs, std::set<std::pair<int, int>>(il));
        }
 
        // Glues the tree with another `rhs` at `indices`. That is, idx in
        // `indices` are considered to be the same vertex. Ids for added
        // vertices are updated accordingly.
        // O(rhs.n).
        value &glue(const value &rhs, std::set<int> indices) {
            std::set<std::pair<int, int>> index_pairs;
            for (auto i : indices)
                index_pairs.emplace(i, i);
            return glue(rhs, index_pairs);
        }
        value &glue(const value &rhs, const std::initializer_list<int> &il) {
            return glue(rhs, std::set<int>(il));
        }
 
        // Prints to std::ostream.
        // O(n).
        friend std::ostream &operator<<(std::ostream &out, const value &val) {
            if (val.print_n_)
                out << val.n() << '\n';
 
            // Prints vertex weights.
            if (val.vertex_weights()) {
                for (int i = 0; i < val.n(); ++i) {
                    if (i > 0)
                        out << " ";
                    out << (*val.vertex_weights())[i];
                }
                out << '\n';
            }
 
            tgen_ensure(static_cast<int>(val.edges().size()) == val.n() - 1,
                        "wtree: value: invalid tree to print (number of edges "
                        "must be `n` - 1)");
 
            // Prints in parent style.
            if (val.print_parents_.has_value()) {
                tgen_ensure(!val.edge_weights().has_value(),
                            "wtree: value: cannot print parent style if edges "
                            "are weighted");
 
                int root = *val.print_parents_;
                bool skip_parent_0 = root == -1;
                if (root == -1)
                    root = 0;
                if (root == val.n())
                    root = next(0, val.n() - 1);
 
                std::vector<int> parent(val.n(), -1);
 
                std::queue<int> q;
                std::vector<int> vis(val.n(), false);
                q.push(root);
                vis[root] = true;
 
                while (q.size()) {
                    int u = q.front();
                    q.pop();
                    for (int v : val.adj()[u])
                        if (!vis[v]) {
                            vis[v] = true;
                            q.push(v);
                            parent[v] = u;
                        }
                }
 
                if (skip_parent_0) {
                    for (int i = 1; i < val.n(); ++i) {
                        tgen_ensure(
                            parent[i] < i,
                            "wtree: value: parent of i must be less than i for "
                            "printing in parent style if root is -1");
 
                        if (i > 1)
                            out << " ";
                        out << parent[i] + val.add_1_;
                    }
                } else {
                    for (int i = 0; i < val.n(); ++i) {
                        if (i > 0)
                            out << " ";
                        out << (parent[i] == -1 ? -1 : parent[i]) + val.add_1_;
                    }
                }
 
                out << '\n';
                return out;
            }
 
            // Prints edges.
            for (int i = 0; i < static_cast<int>(val.edges().size()); ++i) {
                auto [u, v] = val.edges()[i];
                out << (u + val.add_1_) << " " << (v + val.add_1_);
 
                // Edge weight.
                if (val.edge_weights().has_value())
                    out << " " << (*val.edge_weights())[i];
 
                out << '\n';
            }
 
            return out;
        }
 
        // Gets a std::pair<n, adj> representing the value.
        std::pair<int, std::vector<std::set<int>>> to_std() const {
            return std_type(n_, adj_);
        }
 
      private:
        // Adds `k` vertices to the tree (labeled n, n+1, ...n+k-1). Updates
        // `n` accordingly. This makes the tree invalid (not a tree anymore).
        // O(k) amortized.
        value &add_vertices(int k, std::optional<std::vector<VWeight>>
                                       new_vertex_weights = std::nullopt) {
            n_ += k;
            adj_.resize(n());
            if (new_vertex_weights.has_value()) {
                tgen_ensure(vertex_weights().has_value(),
                            "wtree: value: cannot add weighted vertices to "
                            "vertex-unweighted tree");
                tgen_ensure(
                    static_cast<int>(new_vertex_weights->size()) == k,
                    "wtree: value: number of vertex weights must be equal "
                    "to number of added vertices");
 
                vertex_weights_->insert(vertex_weights_->end(),
                                        new_vertex_weights->begin(),
                                        new_vertex_weights->end());
            } else
                tgen_ensure(!vertex_weights().has_value(),
                            "wtree: value: cannot add unweighted vertices to "
                            "vertex-weighted tree");
 
            dsu_.add_elements(k);
 
            return *this;
        }
    };
 
    // Generates tree value.
    // O(n).
    value gen() const {
        // Constructs adjacency list.
        std::vector<std::vector<int>> adj(n_);
        for (auto [u, v] : edges_) {
            adj[u].push_back(v);
            adj[v].push_back(u);
        }
 
        std::vector<int> comp_size;
        std::vector<std::vector<int>> component_ids;
        std::vector<bool> vis(n_, false);
        std::queue<int> q;
 
        for (int i = 0; i < n_; ++i) {
            if (vis[i])
                continue;
 
            vis[i] = true;
            q.push(i);
            comp_size.push_back(0);
            component_ids.emplace_back();
            while (q.size()) {
                int u = q.front();
                q.pop();
                ++comp_size.back();
                component_ids.back().push_back(u);
                for (int v : adj[u]) {
                    if (!vis[v]) {
                        vis[v] = true;
                        q.push(v);
                    }
                }
            }
        }
 
        // Creates edges connecting the connected components by treating them as
        // vertices.
        std::vector<std::pair<int, int>> new_edges(edges_.begin(),
                                                   edges_.end());
        if (comp_size.size() > 1) {
            std::vector<int> prufer_values =
                many_by_distribution(comp_size.size() - 2, comp_size);
            for (auto [u, v] : detail::edges_from_prufer(prufer_values))
                new_edges.emplace_back(pick(component_ids[u]),
                                       pick(component_ids[v]));
        }
 
        return value(n_, new_edges);
    }
 
    // Generates a (not uniformly) random skewed tree.
    // Vertex 0 is the root. For each i in 1 .. n-1, parent(i) is
    // wnext(i, elongation), i.e. a value in [0, i) with skew controlled by
    // elongation (see wnext).
    // If elongation is small enough, generates a star (center 0).
    // If elongation is large enough, generates a path (endpoints 0 and n-1).
    // O(n).
    static value gen_skewed(int n, int elongation) {
        std::vector<std::pair<int, int>> edges;
        for (int i = 1; i < n; ++i)
            edges.emplace_back(i, wnext<int>(i, elongation));
        return value(n, edges);
    }
 
    // Kruskal-like random tree: random vertex pairs until connected.
    // Not uniformly random.
    // O(n log(n) alpha(n)) expected.
    static value gen_kruskal(int n) {
        tgen_ensure(n > 0, "wtree: gen_kruskal: n must be positive");
        if (n == 1)
            return value(1, {});
 
        detail::dsu components(n);
        std::vector<std::pair<int, int>> edges;
        edges.reserve(n - 1);
        while (edges.size() < size_t(n - 1)) {
            int u = next(0, n - 1);
            int v = next(0, n - 1);
            if (u == v)
                continue;
            if (components.unite(u, v))
                edges.emplace_back(u, v);
        }
        return value(n, edges);
    }
};
 
/*
 * Other types of weighted-ness.
 */
 
// Vertex weighted tree.
template <typename VWeight> using vtree = wtree<VWeight, int>;
 
// Edge weighted tree.
template <typename EWeight> using etree = wtree<int, EWeight>;
 
// Unweighted tree.
using tree = wtree<int, int>;
 
/*************
 *           *
 *   GRAPH   *
 *           *
 *************/
 
namespace detail {
 
// Canonical undirected edge key for duplicate detection; stores (min(u, v),
// max(u, v)). O(1).
inline uint64_t undirected_edge_key(int u, int v) {
    if (u > v)
        std::swap(u, v);
    return (static_cast<uint64_t>(u) << 32) |
           static_cast<uint64_t>(static_cast<uint32_t>(v));
}
 
// Directed edge key for duplicate detection; stores (u, v).
// O(1).
inline uint64_t directed_edge_key(int u, int v) {
    return (static_cast<uint64_t>(u) << 32) |
           static_cast<uint64_t>(static_cast<uint32_t>(v));
}
 
// Maximum number of edges in a simple graph on n vertices.
// O(1).
inline long long max_graph_edges(int n, bool directed, bool self_loops) {
    if (n <= 0)
        return 0;
    if (directed)
        return self_loops ? static_cast<long long>(n) * n
                          : static_cast<long long>(n) * (n - 1);
    return self_loops ? static_cast<long long>(n) * (n + 1) / 2
                      : static_cast<long long>(n) * (n - 1) / 2;
}
 
// Uniform random edge for rejection sampling.
// O(1) expected.
inline std::pair<int, int> get_random_graph_edge(int n, bool directed,
                                                 bool self_loops) {
    if (directed) {
        if (self_loops)
            return {next<int>(0, n - 1), next<int>(0, n - 1)};
        int u = next<int>(0, n - 1);
        int v = next<int>(0, n - 1);
        while (u == v)
            v = next<int>(0, n - 1);
        return {u, v};
    }
    if (self_loops) {
        int u = next<int>(0, n - 1);
        int v = next<int>(0, n - 1);
        if (u > v)
            std::swap(u, v);
        return {u, v};
    }
    int u = next<int>(0, n - 1);
    int v = next<int>(0, n - 1);
    while (u == v)
        v = next<int>(0, n - 1);
    if (u > v)
        std::swap(u, v);
    return {u, v};
}
 
// Decodes a linear edge index to (u, v) for an undirected simple graph,
// with u < v.
// O(log n).
inline std::pair<int, int> decode_undirected_simple_edge(int n, long long idx) {
    auto base = [&](int u) -> long long {
        return static_cast<long long>(u) * (n - 1) -
               static_cast<long long>(u) * (u - 1) / 2;
    };
    int lo = 0, hi = n - 2;
    while (lo < hi) {
        int mid = (lo + hi + 1) / 2;
        if (base(mid) <= idx)
            lo = mid;
        else
            hi = mid - 1;
    }
    return {lo, lo + 1 + int(idx - base(lo))};
}
 
// Decodes a linear edge index to (u, v) for an undirected graph with loops,
// with u <= v.
// O(log n).
inline std::pair<int, int> decode_undirected_loops_edge(int n, long long idx) {
    auto base = [&](int u) -> long long {
        return static_cast<long long>(u) * n -
               static_cast<long long>(u) * (u - 1) / 2;
    };
    int lo = 0, hi = n - 1;
    while (lo < hi) {
        int mid = (lo + hi + 1) / 2;
        if (base(mid) <= idx)
            lo = mid;
        else
            hi = mid - 1;
    }
    return {lo, lo + int(idx - base(lo))};
}
 
// Decodes a linear edge index to (u, v) for a directed simple graph (no loops).
// O(1).
inline std::pair<int, int> decode_directed_simple_edge(int n, long long idx) {
    int u = idx / (n - 1);
    int rem = idx % (n - 1);
    return {u, rem + (rem >= u)};
}
 
// Decodes a linear edge index according to graph mode.
// O(log n) for undirected, O(1) for directed.
inline std::pair<int, int>
decode_graph_edge_index(int n, long long idx, bool directed, bool self_loops) {
    if (directed) {
        if (self_loops)
            return {int(idx / n), int(idx % n)};
        return decode_directed_simple_edge(n, idx);
    }
    if (self_loops)
        return decode_undirected_loops_edge(n, idx);
    return decode_undirected_simple_edge(n, idx);
}
 
} // namespace detail
 
/*
 * Graph generator.
 *
 * Graphs of `n` vertices labeled from 0 to n-1 and `m` edges.
 * These are labeled graphs, that is, isomorphism is not taken into
 * account. VWeight is the type of vertex weights, and EWeight is the type of
 * edge weights. Generator does not generate weights. The weights are to be set
 * in the wgraph::value.
 */
 
template <typename VWeight, typename EWeight>
struct wgraph : gen_base<wgraph<VWeight, EWeight>> {
    int n_, m_;                           // Number of vertices and edges.
    std::set<std::pair<int, int>> edges_; // Edges that were set.
    bool is_directed_;                    // If graph is directed.
    bool has_self_loops_;                 // If self-loops are allowed.
 
    // Creates graph generator with `n` vertices and `m` edges.
    // Additionally, you can set if the graph is directed and if self loops are
    // allowed.
    // O(1).
    wgraph(int n, int m, bool is_directed = false, bool has_self_loops = false)
        : n_(n), m_(m), is_directed_(is_directed),
          has_self_loops_(has_self_loops) {
        tgen_ensure(n > 0, "wgraph: number of vertices must be positive");
    }
 
    // Adds edge between u and v (this edge must be generated).
    // O(log m).
    wgraph &add_edge(int u, int v) {
        tgen_ensure(0 <= std::min(u, v) and std::max(u, v) < n_,
                    "wgraph: vertices must be indexed in [0, n)");
 
        if (!is_directed_ and u > v)
            std::swap(u, v);
        edges_.emplace(u, v);
        tgen_ensure(static_cast<int>(edges_.size()) <= m_,
                    "wgraph: too many edges were added");
        return *this;
    }
 
    // Graph value.
    //
    // Edges are stored in both directions (if undirected) in adjacency list,
    // but only u < v in edge list.
    // Optimized for performance (lazy adjacency list; edge-list constructor
    // stores edges only).
    struct value : gen_value_base<value> {
        using std_type = std::tuple<int, int, std::vector<std::set<int>>>;
 
        int n_;                                  // Number of vertices.
        std::vector<std::set<int>> adj_;         // Adjacency list.
        std::vector<std::pair<int, int>> edges_; // Edge list.
        bool is_directed_;                       // If graph is directed.
        bool add_1_;    // If should add 1 for printing vertex ids.
        bool print_nm_; // If should print n and m.
        mutable bool adj_built_{
            false}; // Lazy cache: true once adj_ is built from edges_; mutable
                    // so const adj() can populate it.
        std::optional<std::vector<VWeight>> vertex_weights_; // Vertex weights.
        std::optional<std::vector<EWeight>>
            edge_weights_; // Edge weights (in same order as edges_ ).
 
        // Creates value from adjacency list. The edges
        // are considered to be directed.
        // O(n + m).
        value(const std::vector<std::set<int>> &adj, bool is_directed = false)
            : n_(static_cast<int>(adj.size())), adj_(adj),
              is_directed_(is_directed), add_1_(false), print_nm_(false),
              adj_built_(true) {
            for (int u = 0; u < n_; ++u)
                for (auto v : adj[u]) {
                    tgen_ensure(
                        0 <= v and v < n_,
                        "wgraph: value: vertices must be indexed in [0, n)");
                    // Undirected adjacency is symmetric: count each edge once
                    // (canonical u <= v). Directed: every out-edge appears
                    // once.
                    if (is_directed_ or u <= v)
                        edges_.emplace_back(u, v);
                }
        }
 
        // Creates value from `n`, `m`, and edge list. The edges are
        // considered to be directed.
        // Optimized for performance (lazy adjacency list; unordered_set dedup).
        // O(m log m).
        value(int n, const std::vector<std::pair<int, int>> &edges = {},
              bool is_directed = false)
            : n_(n), edges_(), is_directed_(is_directed), add_1_(false),
              print_nm_(false), adj_built_(false) {
            edges_.reserve(edges.size());
            std::unordered_set<uint64_t> seen;
            seen.reserve(edges.size() * 2 + 1);
            for (auto [u, v] : edges) {
                tgen_ensure(
                    0 <= std::min(u, v) and std::max(u, v) < n,
                    "wgraph: value: vertices must be indexed in [0, n)");
                if (!is_directed_ and u > v)
                    std::swap(u, v);
                uint64_t key = is_directed_ ? detail::directed_edge_key(u, v)
                                            : detail::undirected_edge_key(u, v);
                if (seen.insert(key).second)
                    edges_.emplace_back(u, v);
            }
        }
        value(int n, const std::set<std::pair<int, int>> &edges,
              bool is_directed = false)
            : value(
                  n,
                  std::vector<std::pair<int, int>>(edges.begin(), edges.end()),
                  is_directed) {}
        value(int n, const std::initializer_list<std::pair<int, int>> &edges,
              bool is_directed = false)
            : value(n, std::vector<std::pair<int, int>>(edges), is_directed) {}
 
        // Creates graph from tree (undirected, same edges).
        // O(n).
        value(const typename wtree<VWeight, EWeight>::value &t)
            : value(t.n(), t.edges(), false) {
            if (t.vertex_weights().has_value()) {
                vertex_weights_ = *t.vertex_weights();
            }
            if (t.edge_weights().has_value()) {
                edge_weights_ = *t.edge_weights();
            }
        }
 
        // Weight type conversion.
        // O(n + m).
        template <typename NewVWeight, typename NewEWeight>
        typename wgraph<NewVWeight, NewEWeight>::value
        convert_weight_types() const {
            tgen_ensure(!vertex_weights_.has_value() and
                            !edge_weights_.has_value(),
                        "wgraph: value: cannot convert weight type after "
                        "assigning weights");
 
            ensure_adj_built();
            typename wgraph<NewVWeight, NewEWeight>::value new_graph(
                adj_, is_directed_);
            new_graph.is_directed_ = is_directed_;
            new_graph.add_1_ = add_1_;
            new_graph.print_nm_ = print_nm_;
            return new_graph;
        }
 
        // Fetches number of vertices.
        int n() const { return n_; }
 
        // Fetches number of edges.
        int m() const { return edges_.size(); }
 
        // Fetches if graph is directed;
        bool is_directed() const { return is_directed_; }
 
        // Fetches a const ref. to adjacency list.
        const std::vector<std::set<int>> &adj() const {
            ensure_adj_built();
            return adj_;
        }
 
        // Fetches a const ref. to edge set.
        const std::vector<std::pair<int, int>> &edges() const { return edges_; }
 
        // Fetches vertex weights.
        const std::optional<std::vector<VWeight>> &vertex_weights() const {
            return vertex_weights_;
        }
 
        // Fetches edge weights.
        const std::optional<std::vector<EWeight>> &edge_weights() const {
            return edge_weights_;
        }
 
        // Sets vertex weights.
        // O(n + m).
        template <typename NewVWeight = VWeight>
        typename wgraph<NewVWeight, EWeight>::value set_vertex_weights(
            const std::vector<NewVWeight> &vertex_weights) const {
            tgen_ensure(static_cast<int>(vertex_weights.size()) == n(),
                        "wgraph: value: must give `n` vertex weights");
 
            auto new_graph = convert_weight_types<NewVWeight, EWeight>();
            new_graph.vertex_weights_ = vertex_weights;
            return new_graph;
        }
 
        // Sets edge weights.
        // O(n + m).
        template <typename NewEWeight = EWeight>
        typename wgraph<VWeight, NewEWeight>::value
        set_edge_weights(const std::vector<NewEWeight> &edge_weights) const {
            tgen_ensure(static_cast<int>(edge_weights.size()) == m(),
                        "wgraph: value: must give `m` edge weights");
 
            auto new_graph = convert_weight_types<VWeight, NewEWeight>();
            new_graph.edge_weights_ = edge_weights;
            return new_graph;
        }
 
        // Enables edge-weighted mode before adding weighted edges
        // incrementally. The graph must have no edges yet. O(1).
        value &edge_weighted() {
            tgen_ensure(m() == 0,
                        "wgraph: value: edge_weighted requires a graph with no "
                        "edges");
            tgen_ensure(!edge_weights_.has_value(),
                        "wgraph: value: graph is already edge-weighted");
 
            edge_weights_ = std::vector<EWeight>();
            return *this;
        }
 
        // Adds 1 to vertex ids, for printing.
        // O(1).
        value &add_1() {
            add_1_ = true;
            return *this;
        }
 
        // Prints `n m` on a new line before printing the edges.
        // O(1).
        value &print_nm() {
            print_nm_ = true;
            return *this;
        }
 
        // Shuffles the graph's vertex labels (except those in `indices`,
        // which keep their current label) and edge order. The change is
        // applied eagerly to the underlying adjacency list, edge list,
        // vertex weights and edge weights.
        // O(n + m).
        value &shuffle_except(std::set<int> indices) {
            ensure_adj_built();
            // Builds the relabeling: for each vertex `i`, `new_label[i]` is
            // its new id. Vertices in `indices` keep their label; the others
            // are permuted among themselves.
            std::vector<int> new_label(n());
            std::vector<int> shuffled;
            for (int i = 0; i < n(); ++i) {
                if (indices.count(i))
                    new_label[i] = i;
                else
                    shuffled.push_back(i);
            }
            std::vector<int> targets = shuffled;
            tgen::shuffle(targets.begin(), targets.end());
            for (size_t k = 0; k < shuffled.size(); ++k)
                new_label[shuffled[k]] = targets[k];
 
            // Rewrites adjacency list with new labels.
            std::vector<std::set<int>> new_adj(n());
            for (int u = 0; u < n(); ++u)
                for (int v : adj_[u])
                    new_adj[new_label[u]].insert(new_label[v]);
            adj_ = new_adj;
 
            // Rewrites edges with new labels (canonical undirected order).
            for (auto &[u, v] : edges_) {
                u = new_label[u];
                v = new_label[v];
                if (!is_directed_ and u > v)
                    std::swap(u, v);
            }
 
            // Permutes vertex weights to match the new labels.
            if (vertex_weights_.has_value()) {
                std::vector<VWeight> new_vw(n());
                for (int i = 0; i < n(); ++i)
                    new_vw[new_label[i]] = (*vertex_weights_)[i];
                vertex_weights_ = new_vw;
            }
 
            // Shuffles edge order, keeping edge weights aligned.
 
            std::vector<int> perm(edges_.size());
            std::iota(perm.begin(), perm.end(), 0);
            tgen::shuffle(perm.begin(), perm.end());
 
            std::vector<std::pair<int, int>> new_edges;
            std::optional<std::vector<EWeight>> new_ew;
            if (edge_weights_.has_value())
                new_ew = std::vector<EWeight>();
            for (int i : perm) {
                new_edges.push_back(edges_[i]);
                if (new_ew.has_value())
                    new_ew->push_back((*edge_weights_)[i]);
            }
 
            edges_ = new_edges;
            if (new_ew.has_value())
                edge_weights_ = new_ew;
 
            return *this;
        }
 
        // Shuffles the graph's vertices and edge order.
        // O(n + m).
        value &shuffle() { return shuffle_except({}); }
 
        // Adds `k` vertices to the graph (labeled n, n+1, ...n+k-1). Updates
        // `n` accordingly.
        // O(k) amortized.
        value &add_vertices(int k, std::optional<std::vector<VWeight>>
                                       new_vertex_weights = std::nullopt) {
            ensure_adj_built();
            n_ += k;
            adj_.resize(n());
            if (new_vertex_weights.has_value()) {
                tgen_ensure(vertex_weights().has_value(),
                            "wgraph: value: cannot add weighted vertices to "
                            "vertex-unweighted graph");
                tgen_ensure(
                    static_cast<int>(new_vertex_weights->size()) == k,
                    "wgraph: value: number of vertex weights must be equal "
                    "to number of added vertices");
 
                vertex_weights_->insert(vertex_weights_->end(),
                                        new_vertex_weights->begin(),
                                        new_vertex_weights->end());
            } else
                tgen_ensure(!vertex_weights().has_value(),
                            "wgraph: value: cannot add unweighted vertices to "
                            "vertex-weighted graph");
 
            return *this;
        }
 
        // Adds edge (u, v).
        // O(log n) amortized.
        value &add_edge(int u, int v, std::optional<EWeight> w = std::nullopt) {
            ensure_adj_built();
            tgen_ensure(0 <= std::min(u, v) and std::max(u, v) < n(),
                        "wgraph: value: vertex ids must be valid");
 
            if (!is_directed() and u > v)
                std::swap(u, v);
 
            if (adj_[u].count(v))
                return *this;
 
            adj_[u].insert(v);
            if (!is_directed())
                adj_[v].insert(u);
            edges_.emplace_back(u, v);
 
            if (w.has_value()) {
                tgen_ensure(edge_weights().has_value(),
                            "wgraph: value: cannot add weighted edge to "
                            "edge-unweighted graph");
 
                edge_weights_->push_back(*w);
            } else
                tgen_ensure(!edge_weights().has_value(),
                            "wgraph: value: cannot add unweighted edge to "
                            "edge-weighted graph");
 
            return *this;
        }
 
        // Links graph with another `rhs`, adding the edge between u (in left
        // graph) and v (in right graph). Ids for added vertices are updated
        // accordingly.
        // O(rhs.n + rhs.m * log n) amortized.
        value &link(const value &rhs, int new_u, int new_v,
                    std::optional<EWeight> new_w = std::nullopt) {
            tgen_ensure(0 <= new_u and new_u < n() and 0 <= new_v and
                            new_v < rhs.n(),
                        "wgraph: value: vertex ids must be valid");
 
            // Edges from right-hand side.
            int shift = n();
            add_vertices(rhs.n(), rhs.vertex_weights());
            for (int i = 0; i < rhs.m(); ++i) {
                auto [u, v] = rhs.edges()[i];
                add_edge(shift + u, shift + v,
                         rhs.edge_weights().has_value()
                             ? std::optional<EWeight>((*rhs.edge_weights())[i])
                             : std::nullopt);
            }
 
            // New edge.
            add_edge(new_u, shift + new_v, new_w);
 
            return *this;
        }
 
        // Glues the graph with another `rhs` such that index_pairs[i].first is
        // considered to be the same as index_pairs[i].second. Ids for added
        // vertices are updated accordingly.
        // O(rhs.n + rhs.m * log n) amortized.
        value &glue(const value &rhs,
                    std::set<std::pair<int, int>> index_pairs) {
            tgen_ensure(
                is_directed() == rhs.is_directed(),
                "wgraph: value: graphs must have the same is_directed value");
 
            // Checks validity of indices.
            std::set<int> idx_left, idx_right;
            std::vector<int> right_id_to_left(rhs.n(), -1);
            for (auto [l, r] : index_pairs) {
                tgen_ensure(
                    0 <= l and l < n() and 0 <= r and r < rhs.n(),
                    "wgraph: value: vertex indices to glue must be valid");
                tgen_ensure(idx_left.count(l) == 0 and idx_right.count(r) == 0,
                            "wgraph: value: must not have repeated indices "
                            "on the same side to glue");
 
                idx_left.insert(l);
                idx_right.insert(r);
                right_id_to_left[r] = l;
            }
 
            // Computes new ids of right vertices.
            std::vector<int> new_right_id(rhs.n(), -1);
            int intersection_lt = 0;
            std::optional<std::vector<VWeight>> rhs_vertex_weights;
            for (int i = 0; i < rhs.n(); ++i) {
                if (right_id_to_left[i] != -1) {
                    // Is in intersection.
                    ++intersection_lt;
                    new_right_id[i] = right_id_to_left[i];
                } else {
                    // New id.
                    new_right_id[i] = n() + i - intersection_lt;
                    if (rhs.vertex_weights().has_value()) {
                        if (!rhs_vertex_weights.has_value())
                            rhs_vertex_weights = std::vector<VWeight>();
                        rhs_vertex_weights->push_back(
                            (*rhs.vertex_weights())[i]);
                    }
                }
            }
 
            // Adds new vertices and edges.
            add_vertices(rhs.n() - intersection_lt, rhs_vertex_weights);
            for (int i = 0; i < rhs.m(); ++i) {
                auto [u, v] = rhs.edges()[i];
                add_edge(new_right_id[u], new_right_id[v],
                         rhs.edge_weights().has_value()
                             ? std::optional<EWeight>((*rhs.edge_weights())[i])
                             : std::nullopt);
            }
 
            return *this;
        }
        value &glue(const value &rhs,
                    std::initializer_list<std::pair<int, int>> il) {
            return glue(rhs, std::set<std::pair<int, int>>(il));
        }
 
        // Glues the graph with another `rhs` at `indices`. That is, idx in
        // `indices` are considered to be the same vertex. Ids for added
        // vertices are updated accordingly.
        // O(rhs.n + rhs.m * log n) amortized.
        value &glue(const value &rhs, std::set<int> indices) {
            std::set<std::pair<int, int>> index_pairs;
            for (auto i : indices)
                index_pairs.emplace(i, i);
            return glue(rhs, index_pairs);
        }
        value &glue(const value &rhs, const std::initializer_list<int> &il) {
            return glue(rhs, std::set<int>(il));
        }
 
        // Disjoint union.
        // Shifts ids from `rhs` graph by n().
        // O(rhs.n + rhs.m * log n) amortized.
        value &disjoint_union(const value &rhs) {
            return glue(rhs, std::set<int>());
        }
 
        // Computes uniformly random subgraph of graph with num_edges edges.
        // O(n + m).
        value &random_subgraph(int num_edges) {
            tgen_ensure(
                num_edges <= m(),
                "wgraph: value: can choose at most `m` edges from graph");
 
            std::vector<std::pair<int, int>> new_edges;
            std::optional<std::vector<EWeight>> new_edge_weights;
 
            int left = m();
            for (int i = 0; i < m(); ++i) {
                if (next(1, left--) <= num_edges) {
                    new_edges.push_back(edges()[i]);
                    if (edge_weights_.has_value()) {
                        if (!new_edge_weights.has_value())
                            new_edge_weights = std::vector<EWeight>();
                        new_edge_weights->push_back((*edge_weights())[i]);
                    }
                    --num_edges;
                }
            }
 
            edges_ = new_edges;
            edge_weights_ = new_edge_weights;
            rebuild_adj_from_edge_list();
            return *this;
        }
 
        // Computes a random (not uniform) subgraph with `num_edges` edges that
        // keeps every connected component connected (does not increase the
        // number of connected components).
        // 1. Picks a spanning forest via randomized Prim.
        // 2. Adds additional edges uniformly at random.
        // O(n + m).
        value &random_connected_subgraph(int num_edges) {
            tgen_ensure(!is_directed_,
                        "wgraph: value: random_connected_subgraph is only for "
                        "undirected graphs");
            tgen_ensure(
                num_edges <= m(),
                "wgraph: value: can choose at most `m` edges from graph");
 
            // Builds an incidence list: for each vertex, the (neighbor, edge
            // index) pairs.
            std::vector<std::vector<std::pair<int, int>>> incident(n());
            for (int i = 0; i < m(); ++i) {
                auto [u, v] = edges_[i];
                incident[u].emplace_back(v, i);
                incident[v].emplace_back(u, i);
            }
 
            // Randomized Prim.
            std::vector<bool> vis(n(), false);
            std::vector<int> queue;
            std::vector<bool> in_tree(m(), false);
            int forest_edges = 0;
 
            for (int start = 0; start < n(); ++start) {
                if (vis[start])
                    continue;
                vis[start] = true;
                queue.push_back(start);
 
                while (!queue.empty()) {
                    int i = tgen::next<int>(0, queue.size() - 1);
                    int u = queue[i];
                    std::swap(queue[i], queue.back());
                    queue.pop_back();
 
                    for (auto [v, edge_idx] : incident[u]) {
                        if (!vis[v]) {
                            vis[v] = true;
                            queue.push_back(v);
                            in_tree[edge_idx] = true;
                            ++forest_edges;
                        }
                    }
                }
            }
            tgen_ensure(
                num_edges >= forest_edges,
                "wgraph: value: random_connected_subgraph needs at least "
                "`n - c` edges, where `c` is the number of connected "
                "components");
 
            // Splits edge indices into forest edges and the rest.
            std::vector<int> tree_idx, rest_idx;
            for (int i = 0; i < m(); ++i) {
                if (in_tree[i])
                    tree_idx.push_back(i);
                else
                    rest_idx.push_back(i);
            }
 
            tgen::shuffle(rest_idx.begin(), rest_idx.end());
 
            std::vector<int> chosen_idx;
            chosen_idx.insert(chosen_idx.end(), tree_idx.begin(),
                              tree_idx.end());
            chosen_idx.insert(chosen_idx.end(), rest_idx.begin(),
                              rest_idx.begin() + num_edges - forest_edges);
 
            detail::tgen_ensure_against_bug(
                static_cast<int>(chosen_idx.size()) == num_edges,
                "wgraph: value: chose a wrong number of edges");
 
            std::vector<std::pair<int, int>> new_edges;
            std::optional<std::vector<EWeight>> new_edge_weights;
            if (edge_weights_.has_value())
                new_edge_weights = std::vector<EWeight>();
            for (int i : chosen_idx) {
                new_edges.push_back(edges_[i]);
                if (new_edge_weights.has_value())
                    new_edge_weights->push_back((*edge_weights_)[i]);
            }
 
            edges_ = new_edges;
            edge_weights_ = new_edge_weights;
            rebuild_adj_from_edge_list();
            return *this;
        }
 
        // Complement. Self loops are maintained.
        // O(n^2).
        value operator!() const {
            tgen_ensure(!edge_weights_.has_value(),
                        "wgraph: value: cannot compute complement of "
                        "edge-weighted graph");
 
            value complement = *this;
            complement.ensure_adj_built();
            std::vector<std::pair<int, int>> compl_edges;
            for (int i = 0; i < complement.n_; ++i) {
                std::set<int> complement_adj;
                for (int j = 0; j < complement.n_; ++j) {
                    bool add_j = false;
                    if (j == i and complement.adj_[i].count(j))
                        add_j = true;
                    if (j != i and !complement.adj_[i].count(j))
                        add_j = true;
 
                    if (add_j) {
                        complement_adj.insert(j);
                        // If i > j and !is_directed(), we don't add the edge.
                        if (i <= j or complement.is_directed_) {
                            compl_edges.emplace_back(i, j);
                        }
                    }
                }
                std::swap(complement.adj_[i], complement_adj);
            }
            std::swap(complement.edges_, compl_edges);
 
            return complement;
        }
 
        // Concatenates two values.
        // O(N + M log N), N = n + rhs.n, M = m + rhs.m.
        value operator+(const value &rhs) const {
            tgen_ensure(is_directed() == rhs.is_directed(),
                        "wgraph: value: graphs must have the same "
                        "is_directed value");
 
            tgen_ensure(vertex_weights().has_value() ==
                            rhs.vertex_weights().has_value(),
                        "wgraph: value: cannot concatenate vertex-weighted "
                        "wgraph to unweighted");
            tgen_ensure(edge_weights().has_value() ==
                            rhs.edge_weights().has_value(),
                        "wgraph: value: cannot concatenate edge-weighted "
                        "wgraph to unweighted");
 
            value concat = *this;
            concat.glue(rhs, std::set<std::pair<int, int>>());
            concat.add_1_ = add_1_ | rhs.add_1_;
            concat.print_nm_ = print_nm_ | rhs.print_nm_;
 
            return concat;
        }
 
        // Prints to std::ostream.
        // O(n + m).
        friend std::ostream &operator<<(std::ostream &out, const value &val) {
            // Prints `n` and `m`.
            if (val.print_nm_)
                out << val.n() << " " << val.m() << '\n';
 
            // Prints vertex weights.
            if (val.vertex_weights()) {
                for (int i = 0; i < val.n(); ++i) {
                    if (i > 0)
                        out << " ";
                    out << (*val.vertex_weights())[i];
                }
                out << '\n';
            }
 
            // Prints edges.
            for (int i = 0; i < val.m(); ++i) {
                auto [u, v] = val.edges()[i];
                out << (u + val.add_1_) << " " << (v + val.add_1_);
 
                // Edge weight.
                if (val.edge_weights().has_value())
                    out << " " << (*val.edge_weights())[i];
 
                out << '\n';
            }
 
            return out;
        }
 
        // Gets a std::tuple<n, m, adj> representing the value.
        std::tuple<int, int, std::vector<std::set<int>>> to_std() const {
            ensure_adj_built();
            return std_type(n_, m(), adj_);
        }
 
      private:
        // Rebuilds adjacency from edges_ after replacing the edge list (e.g.
        // subgraph operations).
        // O(m log n).
        void rebuild_adj_from_edge_list() {
            adj_.assign(n_, {});
            for (auto [u, v] : edges_) {
                adj_[u].insert(v);
                if (!is_directed_)
                    adj_[v].insert(u);
            }
            adj_built_ = true;
        }
 
        // Builds adj_ from edges_ on first use.
        // O(1) if already built; O(m log n) otherwise.
        void ensure_adj_built() const {
            if (adj_built_)
                return;
            const_cast<value *>(this)->rebuild_adj_from_edge_list();
        }
    };
 
    // Adds all edges from `rhs` as preset edges.
    // O(rhs.m * log m).
    wgraph &add_edges_from(const value &rhs) {
        tgen_ensure(is_directed_ == rhs.is_directed(),
                    "wgraph: graphs must have the same is_directed value");
 
        for (auto [u, v] : rhs.edges())
            add_edge(u, v);
        return *this;
    }
 
    // Generates graph value.
    // Optimized for performance: dense no-preset graphs use index sampling;
    // otherwise gen_remaining_edges.
    // O(n + m log^2 n) expected.
    value gen() const {
        detail::tgen_ensure_against_bug(static_cast<int>(edges_.size()) <= m_,
                                        "wgraph: too many edges were added");
 
        // All edges already added.
        if (static_cast<int>(edges_.size()) == m_)
            return value(n_, edges_, is_directed_);
 
        // Splits into two cases to optimize performance.
 
        // No presets and m > max_edges / 2: sample m distinct edge indices.
        if (auto indexed = try_gen_by_edge_index())
            return *indexed;
 
        // Otherwise: fill preset edges up to m_ with uniform random edges.
        return gen_remaining_edges(
            std::vector<std::pair<int, int>>(edges_.begin(), edges_.end()));
    }
 
    // Gets a (not uniformly) random connected undirected graph.
    // 1. Preset edges induce a spanning forest on their components.
    // 2. Then, uniformly random edges between components are added.
    // 3. Remaining edges are added uniformly at random.
    // O(n + m log^2 n) expected.
    value get_connected() const {
        tgen_ensure(!is_directed_,
                    "wgraph: get_connected is only for undirected graphs");
        tgen_ensure(m_ >= n_ - 1,
                    "wgraph: connected graph needs at least n - 1 edges");
 
        std::vector<std::pair<int, int>> edges;
        edges.reserve(m_);
 
        if (edges_.empty()) {
            if (n_ > 1) {
                std::vector<int> prufer(n_ - 2);
                for (int i = 0; i < n_ - 2; ++i)
                    prufer[i] = next<int>(0, n_ - 1);
                for (auto [u, v] : detail::edges_from_prufer(std::move(prufer)))
                    edges.emplace_back(u, v);
            }
        } else {
            edges.assign(edges_.begin(), edges_.end());
 
            std::vector<std::vector<int>> adj(n_);
            for (auto [u, v] : edges_) {
                adj[u].push_back(v);
                adj[v].push_back(u);
            }
 
            std::vector<int> comp_size;
            std::vector<std::vector<int>> component_ids;
            std::vector<bool> vis(n_, false);
            std::queue<int> q;
 
            for (int i = 0; i < n_; ++i) {
                if (vis[i])
                    continue;
 
                vis[i] = true;
                q.push(i);
                comp_size.push_back(0);
                component_ids.emplace_back();
                while (q.size()) {
                    int u = q.front();
                    q.pop();
                    ++comp_size.back();
                    component_ids.back().push_back(u);
                    for (int v : adj[u]) {
                        if (!vis[v]) {
                            vis[v] = true;
                            q.push(v);
                        }
                    }
                }
            }
 
            if (component_ids.size() > 1) {
                std::vector<int> prufer_values =
                    many_by_distribution(component_ids.size() - 2, comp_size);
                for (auto [u, v] :
                     detail::edges_from_prufer(std::move(prufer_values)))
                    edges.emplace_back(pick(component_ids[u]),
                                       pick(component_ids[v]));
            }
        }
 
        return gen_remaining_edges(std::move(edges));
    }
 
    // Gets a (not uniformly) random directed acyclic graph.
    // 1. Randomized Kahn (uniform choice among indegree-0 vertices) yields a
    //    random topological order of the preset edges (which must be acyclic).
    // 2. Extra edges are sampled randomly using the order.
    // With no preset edges: sample a random graph then orient acyclically.
    // Optimized for performance (distinct upper-triangle edge-index sampling;
    // rejection instead of pair::distinct for preset edges).
    // O(n + m log^2 n) expected.
    value get_acyclic() const {
        tgen_ensure(is_directed_,
                    "wgraph: get_acyclic is only for directed graphs");
 
        if (edges_.empty()) {
            std::vector<int> order(n_);
            std::iota(order.begin(), order.end(), 0);
            for (int i = n_ - 1; i > 0; --i)
                std::swap(order[i], order[next(0, i)]);
 
            const long long max_pairs =
                static_cast<long long>(n_) * (n_ - 1) / 2;
            tgen_ensure(m_ <= max_pairs,
                        "wgraph: not enough edges to generate");
 
            std::vector<std::pair<int, int>> edges;
            edges.reserve(m_);
            for (long long idx : distinct_range<long long>(0, max_pairs - 1)
                                     .gen_list(m_)
                                     .to_std()) {
                auto [i, j] = detail::decode_undirected_simple_edge(n_, idx);
                edges.emplace_back(order[i], order[j]);
            }
            return value(n_, edges, true);
        }
 
        std::vector<std::vector<int>> adj(n_);
        std::vector<int> indeg(n_, 0);
        for (auto [u, v] : edges_) {
            adj[u].push_back(v);
            ++indeg[v];
        }
 
        std::vector<int> available;
        for (int i = 0; i < n_; ++i)
            if (indeg[i] == 0)
                available.push_back(i);
 
        // Random topological order using randomized Kahn's algorithm.
        std::vector<int> order;
        while (!available.empty()) {
            int idx = next(0, static_cast<int>(available.size()) - 1);
            int u = available[idx];
            std::swap(available[idx], available.back());
            available.pop_back();
 
            order.push_back(u);
            for (int v : adj[u])
                if (--indeg[v] == 0)
                    available.push_back(v);
        }
 
        tgen_ensure(static_cast<int>(order.size()) == n_,
                    "wgraph: preset edges contain a directed cycle");
 
        value acyclic(n_, edges_, true);
 
        // Generates final edges.
 
        detail::tgen_ensure_against_bug(acyclic.m() <= m_,
                                        "wgraph: too many edges were added");
 
        if (acyclic.m() < m_) {
            std::vector<int> order_pos(n_);
            for (int i = 0; i < n_; ++i)
                order_pos[order[i]] = i;
 
            std::unordered_set<uint64_t> seen;
            seen.reserve(m_ * 2);
            for (auto [u, v] : acyclic.edges())
                seen.insert(
                    detail::undirected_edge_key(order_pos[u], order_pos[v]));
 
            const long long max_pairs =
                static_cast<long long>(n_) * (n_ - 1) / 2;
            while (acyclic.m() < m_) {
                std::pair<int, int> edge;
                if (!detail::try_generate_distinct(seen, [&] {
                        long long idx = next<long long>(0, max_pairs - 1);
                        edge = detail::decode_undirected_simple_edge(n_, idx);
                        return detail::undirected_edge_key(edge.first,
                                                           edge.second);
                    }))
                    throw detail::error("wgraph: not enough edges to generate");
                acyclic.add_edge(order[edge.first], order[edge.second]);
            }
        }
 
        return acyclic;
    }
 
    // Generates a (not uniformly) random skewed connected graph.
    // 1. Builds the same skewed labeled tree as wtree::gen_skewed(n,
    //    elongation)(root 0, parent(i) = wnext(i, elongation) for i >= 1).
    //    If is_directed, tree edges are oriented down the tree.
    // 2. Adds the remaining edges: pick an endpoint u uniformly;
    //    pick k uniformly in [1, spread]; walk from u toward the root k
    //    times along tree parents to get v; add edge (v, u).
    // If elongation is small, generates a graph with small diameter.
    // If elongation is large, generates a graph with large diameter, with
    // vertices 0 and n-1 being far apart.
    // O(n + m log n) if spread is O(1);
    // O(n log n + m log^2 n) expected otherwise.
    static value gen_skewed(int n, int m, int elongation, int spread,
                            bool is_directed = false) {
        tgen_ensure(
            m >= n - 1,
            "wgraph: skewed graph needs at least n - 1 edges to be connected");
        tgen_ensure(spread >= 2,
                    "wgraph: gen_skewed spread must be at least 2");
 
        value skewed(n, {}, is_directed);
 
        std::vector<int> parent(n), depth(n, 0);
        parent[0] = 0;
        for (int i = 1; i < n; ++i) {
            int p = wnext<int>(i, elongation);
            parent[i] = p;
            depth[i] = depth[p] + 1;
            skewed.add_edge(p, i);
        }
 
        const int extra = m - (n - 1);
        if (extra == 0)
            return skewed;
 
        // If spread is large, use binary lifting to find the ancestor.
        // Otherwise, enumerate O(n * spread) ancestor edges and sample
        // directly.
        constexpr int naive_ancestor_spread = 20;
 
        if (spread <= naive_ancestor_spread) {
            std::vector<std::pair<int, int>> candidates;
            candidates.reserve(n * spread);
            for (int u = 0; u < n; ++u) {
                int max_k = std::min(spread, depth[u]);
                if (max_k < 2)
                    continue;
                int v = parent[u];
                for (int k = 2; k <= max_k; ++k) {
                    v = parent[v];
                    candidates.emplace_back(v, u);
                }
            }
 
            tgen_ensure(extra <= static_cast<int>(candidates.size()),
                        "wgraph: not enough edges to generate");
 
            for (auto [v, u] : choose(candidates, extra))
                skewed.add_edge(v, u);
        } else {
            // Binary lifting.
            int lg = 1;
            while ((1 << lg) <= n)
                ++lg;
 
            std::vector<std::vector<int>> up(lg, std::vector<int>(n));
            for (int v = 0; v < n; ++v)
                up[0][v] = parent[v];
            for (int j = 1; j < lg; ++j)
                for (int v = 0; v < n; ++v)
                    up[j][v] = up[j - 1][up[j - 1][v]];
 
            // Creates uniform generator of edges (u, v) such that v is ancestor
            // of u. For that, every u has depth[u]-1 choices for v, so we
            // weight u by min(spread - 1, depth[u] - 1). After that we can
            // just pick the ancestor uniformly.
            std::vector<int> distribution = depth;
            for (int &d : distribution)
                d = std::max(0, std::min(spread - 1, d - 1));
            weighted_sampler vertex_choice(distribution);
            distinct extra_edges([&]() -> std::pair<int, int> {
                int u = vertex_choice.next();
                int k = next(2, spread);
                int v = u;
                for (int j = 0; j < lg; ++j)
                    if (k >> j & 1)
                        v = up[j][v];
                return {v, u};
            });
 
            while (skewed.m() < m) {
                std::pair<int, int> edge;
                try {
                    edge = extra_edges.gen();
                } catch (const std::runtime_error &e) {
                    if (std::string(e.what()) ==
                        "tgen: distinct: no more distinct values")
                        throw detail::error(
                            "wgraph: not enough edges to generate");
                    throw e;
                }
 
                skewed.add_edge(edge.first, edge.second);
            }
        }
 
        return skewed;
    }
 
    // Generates a random bipartite graph. The first side has vertices
    // 0 .. n1-1, the second n1 .. n1+n2-1.
    // Uniform when connected is false (distinct cross-edge indices).
    // When connected, bipartite Prüfer + rejection fill; not uniform over
    // connected bipartite graphs.
    // O(n1 + n2 + m log(n1 * n2)) expected.
    static value gen_bipartite(int n1, int n2, int m, bool connected = false) {
        tgen_ensure(m >= 0, "wgraph: number of edges must be nonnegative");
        long long num_edges = 1LL * n1 * n2;
        tgen_ensure(m <= num_edges,
                    "wgraph: bipartite graph has at most n1 * n2 edges");
        if (connected)
            tgen_ensure(
                m >= n1 + n2 - 1,
                "wgraph: connected bipartite graph needs at least n1 + n2 - 1 "
                "edges");
 
        if (!connected) {
            std::vector<std::pair<int, int>> edges;
            edges.reserve(m);
            for (long long idx : distinct_range<long long>(0, num_edges - 1)
                                     .gen_list(m)
                                     .to_std())
                edges.emplace_back(static_cast<int>(idx / n2),
                                   n1 + static_cast<int>(idx % n2));
            return value(n1 + n2, std::move(edges), false);
        }
 
        std::unordered_set<uint64_t> used_edges;
        used_edges.reserve(m * 2);
        std::vector<std::pair<int, int>> edges;
        edges.reserve(m);
 
        auto pack_edge = [](int u, int v) -> uint64_t {
            if (u > v)
                std::swap(u, v);
            return (static_cast<uint64_t>(u) << 32) | static_cast<uint32_t>(v);
        };
 
        if (n1 > 0 and n2 > 0) {
            std::vector<int> prufer(n1 + n2 - 2);
            for (int i = 0; i < n2 - 1; ++i)
                prufer[i] = next(0, n1 - 1);
            for (int i = 0; i < n1 - 1; ++i)
                prufer[n2 - 1 + i] = next(n1, n1 + n2 - 1);
            shuffle(prufer.begin(), prufer.end());
            for (auto [u, v] : detail::edges_from_prufer(std::move(prufer))) {
                if (u > v)
                    std::swap(u, v);
                if (used_edges.insert(pack_edge(u, v)).second)
                    edges.emplace_back(u, v);
            }
            detail::tgen_ensure_against_bug(
                used_edges.size() == size_t(n1 + n2 - 1),
                "wgraph: invalid bipartite spanning tree size");
        }
 
        while (edges.size() < size_t(m)) {
            int u = next(0, n1 - 1);
            int v = next(n1, n1 + n2 - 1);
            if (used_edges.insert(pack_edge(u, v)).second)
                edges.emplace_back(u, v);
        }
 
        return value(n1 + n2, std::move(edges), false);
    }
 
  private:
    // If this generator has no preset edges and m is large relative to the
    // maximum edge count, sample by distinct edge index. Otherwise
    // std::nullopt.
    // Optimized for performance (index sampling instead of rejection).
    // O(m log n).
    std::optional<value> try_gen_by_edge_index() const {
        if (!edges_.empty())
            return std::nullopt;
 
        long long max_edges =
            detail::max_graph_edges(n_, is_directed_, has_self_loops_);
        if (m_ > max_edges)
            throw detail::error("wgraph: not enough edges to generate");
        if (max_edges <= 0 or 2LL * m_ <= max_edges)
            return std::nullopt;
 
        std::vector<std::pair<int, int>> edges;
        edges.reserve(m_);
        for (long long idx :
             distinct_range<long long>(0, max_edges - 1).gen_list(m_).to_std())
            edges.push_back(detail::decode_graph_edge_index(
                n_, idx, is_directed_, has_self_loops_));
 
        return value(n_, edges, is_directed_);
    }
 
    // Fills `edges` up to m_ with uniform random edges not already present.
    // Optimized for performance (uint64 edge keys + try_generate_distinct).
    // O(m log^2 n) expected.
    value gen_remaining_edges(std::vector<std::pair<int, int>> edges) const {
        detail::tgen_ensure_against_bug(static_cast<int>(edges.size()) <= m_,
                                        "wgraph: too many edges were added");
 
        if (static_cast<int>(edges.size()) == m_)
            return value(n_, edges, is_directed_);
 
        edges.reserve(m_);
 
        std::unordered_set<uint64_t> seen;
        seen.reserve(m_ * 2);
        for (auto [u, v] : edges) {
            if (!is_directed_ and u > v)
                std::swap(u, v);
            seen.insert(is_directed_ ? detail::directed_edge_key(u, v)
                                     : detail::undirected_edge_key(u, v));
        }
 
        while (static_cast<int>(edges.size()) < m_) {
            std::pair<int, int> edge;
            if (!detail::try_generate_distinct(seen, [&] {
                    edge = detail::get_random_graph_edge(n_, is_directed_,
                                                         has_self_loops_);
                    if (!is_directed_ and edge.first > edge.second)
                        std::swap(edge.first, edge.second);
                    return is_directed_ ? detail::directed_edge_key(edge.first,
                                                                    edge.second)
                                        : detail::undirected_edge_key(
                                              edge.first, edge.second);
                }))
                throw detail::error("wgraph: not enough edges to generate");
            edges.emplace_back(edge);
        }
 
        return value(n_, edges, is_directed_);
    }
};
 
// Implementation of wtree::value constructor from wgraph.
// O(n + m alpha(n)).
template <typename VWeight, typename EWeight>
wtree<VWeight, EWeight>::value::value(
    const typename wgraph<VWeight, EWeight>::value &g)
    : n_(g.n()), adj_(g.n()), add_1_(false), print_n_(false), dsu_(g.n()) {
    tgen_ensure(g.n() > 0, "wtree: value: graph must have at least one vertex");
    tgen_ensure(!g.is_directed(),
                "wtree: value: graph must be undirected to form a tree");
 
    if (g.vertex_weights().has_value())
        vertex_weights_ = *g.vertex_weights();
    if (g.edge_weights().has_value())
        edge_weights_ = std::vector<EWeight>();
 
    if (n_ == 1)
        return;
 
    std::vector<int> order(g.m());
    std::iota(order.begin(), order.end(), 0);
    tgen::shuffle(order.begin(), order.end());
 
    std::vector<std::pair<int, int>> tree_edges;
    tree_edges.reserve(n_ - 1);
 
    for (int i : order) {
        auto [u, v] = g.edges()[i];
        if (!dsu_.unite(u, v))
            continue;
        if (u > v)
            std::swap(u, v);
 
        tree_edges.emplace_back(u, v);
        adj_[u].insert(v);
        adj_[v].insert(u);
        if (edge_weights_.has_value())
            edge_weights_->push_back((*g.edge_weights())[i]);
        if (static_cast<int>(tree_edges.size()) == n_ - 1)
            break;
    }
 
    tgen_ensure(static_cast<int>(tree_edges.size()) == n_ - 1,
                "wtree: value: graph must be connected to form a tree");
 
    edges_ = std::move(tree_edges);
}
 
/*
 * Other types of weighted-ness.
 */
 
// Vertex weighted graph.
template <typename VWeight> using vgraph = wgraph<VWeight, int>;
 
// Edge weighted graph.
template <typename EWeight> using egraph = wgraph<int, EWeight>;
 
// Unweighted graph.
using graph = wgraph<int, int>;
 
/*
 * Standard graphs.
 */
 
// Complete.
// O(n^2).
inline graph::value K(int n) { return graph(n, n * (n - 1) / 2).gen(); }
 
// Path.
// Path with `n` vertices. The edges of the path are 0 and n-1.
// If directed, edges are i -> i+1 for i in [0, n-2).
// O(n).
inline graph::value P(int n, bool is_directed = false) {
    graph g(n, n - 1, is_directed);
    for (int i = 0; i + 1 < n; ++i)
        g.add_edge(i, i + 1);
    return g.gen();
}
 
// Cycle.
// n >= 3.
// If directed, edges are i -> (i+1) % n.
// O(n).
inline graph::value C(int n, bool is_directed = false) {
    tgen_ensure(n >= 3, "graph: cycle size must be at least 3");
 
    graph g(n, n, is_directed);
    for (int i = 0; i < n; ++i)
        g.add_edge(i, (i + 1) % n);
    return g.gen();
}
 
// Complete bipartite.
// The first side has vertices `0` to `n1-1`, the second side has vertices `n1`
// to `n1+n2-1`.
// O(n1 * n2).
inline graph::value K(int n1, int n2) {
    graph g(n1 + n2, static_cast<long long>(n1) * n2);
    for (int i = 0; i < n1; ++i)
        for (int j = 0; j < n2; ++j)
            g.add_edge(i, n1 + j);
    return g.gen();
}
 
// Star.
// The center is vertex 0.
// O(n).
inline graph::value S(int n) { return K(1, n - 1); }
 
/****************
 *              *
 *   GEOMETRY   *
 *              *
 ****************/
 
namespace geometry {
 
// Point on the plane with coordinates of type T.
template <typename T> struct point {
    static_assert(std::is_arithmetic_v<T>,
                  "point requires an arithmetic coordinate type");
 
    // Dot/cross product type: __int128 for T = long long, long long for other
    // integral T, T for floating-point.
    using product_t = std::conditional_t<
        std::is_same_v<T, long long>, detail::i128,
        std::conditional_t<std::is_integral_v<T>, long long, T>>;
 
    // x and y coordinates.
    T x_, y_;
 
    // Constructs a point with coordinates x and y.
    point(T x = 0, T y = 0) : x_(x), y_(y) {}
 
    // Returns the x coordinate.
    T x() const { return x_; }
 
    // Returns the y coordinate.
    T y() const { return y_; }
 
    // Equality of coordinates, with epsilon-based equality for floating-point
    // coordinates (tolerance 1e-9).
    static bool coord_eq(T a, T b) {
        if constexpr (std::is_integral_v<T>)
            return a == b;
        constexpr T eps = T(1e-9);
        T d = a - b;
        return d >= -eps and d <= eps;
    }
 
    // Lexicographic order (by x, then y).
    bool operator<(const point &p) const {
        if (!coord_eq(x_, p.x()))
            return x_ < p.x();
        return y_ < p.y();
    }
 
    // Equality of coordinates.
    bool operator==(const point &p) const {
        return coord_eq(x_, p.x()) and coord_eq(y_, p.y());
    }
 
    // Vector addition.
    point operator+(const point &p) const {
        return point(x_ + p.x(), y_ + p.y());
    }
 
    // Vector subtraction.
    point operator-(const point &p) const {
        return point(x_ - p.x(), y_ - p.y());
    }
 
    // Scalar multiplication.
    point operator*(T c) const { return point(x_ * c, y_ * c); }
 
    // Dot product.
    product_t operator*(const point &p) const {
        if constexpr (std::is_floating_point_v<T>)
            return x_ * p.x() + y_ * p.y();
        return product_t(x_) * p.x() + product_t(y_) * p.y();
    }
 
    // Cross product (signed area of the parallelogram).
    product_t operator^(const point &p) const {
        if constexpr (std::is_floating_point_v<T>)
            return x_ * p.y() - y_ * p.x();
        return product_t(x_) * p.y() - product_t(y_) * p.x();
    }
 
    // Prints the point as "x y".
    friend std::ostream &operator<<(std::ostream &out, const point &p) {
        return out << p.x() << ' ' << p.y();
    }
};
 
// Generates n distinct integer points in [min_coord, max_coord]^2 with no three
// collinear.
// O(n).
inline std::vector<point<long long>>
random_points_general_position(int n, long long min_coord,
                               long long max_coord) {
    tgen_ensure(n > 0,
                "geometry: random_points_general_position: n must be positive");
    tgen_ensure(max_coord >= min_coord,
                "geometry: random_points_general_position: min_coord must be "
                "at most max_coord");
    tgen_ensure(
        static_cast<detail::i128>(max_coord) - min_coord <=
            std::numeric_limits<long long>::max(),
        "geometry: random_points_general_position: coordinate range too large");
    uint64_t width = max_coord - min_coord;
    uint64_t p = math::prime_from(2 * n);
 
    // Requires width >= p - 1 because sheared coordinates lie in [0, p - 1].
    tgen_ensure(width >= p - 1,
                "geometry: random_points_general_position: coordinate range "
                "too small for n");
 
    // Base set: (x, x^-1 mod p) for x = 1, ..., p - 1.
    //
    // For a line ax + by + c = 0, substituting y = x^-1 gives ax^2 + cx + b = 0
    // (for x != 0), a quadratic with at most two roots in F_p. So at most two
    // base points lie on any line. x |-> x^-1 is bijective on {1, ..., p - 1},
    // so all points are distinct and no three are collinear.
    std::vector<uint64_t> x_range(p - 1);
    std::iota(x_range.begin(), x_range.end(), 1);
    shuffle(x_range.begin(), x_range.end());
    std::vector<detail::i128> bx(n), by(n);
    for (int i = 0; i < n; ++i) {
        uint64_t x = x_range[i];
        bx[i] = x;
        by[i] = math::modular_inverse(x, p);
    }
 
    // Randomize placement without breaking general position: compose elementary
    // shears in SL(2, F_p), each either [1 r; 0 1] or [1 0; r 1] with
    // r in {-2, -1, 1, 2} (mod p). Every shear has determinant 1, so their
    // product is invertible. Invertible linear maps preserve collinearity, so
    // the image still has no three collinear points.
    const int num_shears = 8;
    std::vector<detail::i128> lin_x = bx, lin_y = by;
 
    for (int it = 0; it < num_shears; ++it) {
        bool vertical_shear = next(2) == 0;
        int shear_r = pick({-2, -1, 1, 2});
 
        for (int i = 0; i < n; ++i) {
            if (vertical_shear)
                lin_x[i] = (lin_x[i] + shear_r * lin_y[i]) % p;
            else
                lin_y[i] = (lin_y[i] + shear_r * lin_x[i]) % p;
 
            if (lin_x[i] < 0)
                lin_x[i] += p;
            if (lin_y[i] < 0)
                lin_y[i] += p;
        }
    }
 
    detail::i128 min_x = lin_x[0], max_x = lin_x[0], min_y = lin_y[0],
                 max_y = lin_y[0];
    for (int i = 1; i < n; ++i) {
        min_x = std::min(min_x, lin_x[i]);
        max_x = std::max(max_x, lin_x[i]);
        min_y = std::min(min_y, lin_y[i]);
        max_y = std::max(max_y, lin_y[i]);
    }
 
    long long x_shift =
        min_coord - min_x + next<long long>(0, width - (max_x - min_x));
    long long y_shift =
        min_coord - min_y + next<long long>(0, width - (max_y - min_y));
 
    std::vector<point<long long>> pts;
    for (int i = 0; i < n; ++i)
        pts.emplace_back(lin_x[i] + x_shift, lin_y[i] + y_shift);
    return pts;
}
 
namespace detail {
 
using i128 = tgen::detail::i128;
 
// Signed area of triangle (a, b, p); positive iff (a, b, p) are in
// counterclockwise order. 0 iff (a, b, p) are collinear. O(1).
inline i128 ccw(const point<long long> &a, const point<long long> &b,
                const point<long long> &p) {
    return (static_cast<i128>(b.x()) - a.x()) *
               (static_cast<i128>(p.y()) - a.y()) -
           (static_cast<i128>(b.y()) - a.y()) *
               (static_cast<i128>(p.x()) - a.x());
}
 
// Integer projection of P onto line AB (A and B need not be distinct).
inline i128 proj_on_ab(const point<long long> &P, const point<long long> &A,
                       const point<long long> &B) {
    return (P - A) * (B - A);
}
 
// In-place Hamiltonian path on points[left..right-1] with points[left]
// start and points[right-1] end.
// O(n log n) expected if points are "random", O(n^2) worst case.
inline void conquer(std::vector<point<long long>> &points, int left,
                    int right) {
    if (right - left <= 3)
        return;
 
    point<long long> A = points[left], B = points[right - 1];
 
    // If all points are collinear, sort them properly and return.
    bool all_collinear = true;
    for (int k = left + 1; k < right - 1; ++k) {
        if (ccw(A, B, points[k]) != 0) {
            all_collinear = false;
            break;
        }
    }
    if (all_collinear) {
        std::sort(points.begin() + left, points.begin() + right,
                  [&](const point<long long> &P, const point<long long> &Q) {
                      return proj_on_ab(P, A, B) < proj_on_ab(Q, A, B);
                  });
        return;
    }
 
    // Choses a pivot that is not collinear with A and B.
    std::vector<int> candidates;
    for (int k = left + 1; k < right - 1; ++k) {
        if (ccw(A, B, points[k]) != 0)
            candidates.push_back(k);
    }
    int ci = candidates[next(0, static_cast<int>(candidates.size()) - 1)];
    point<long long> C = points[ci];
 
    uint64_t wa = next<uint64_t>(1, std::numeric_limits<uint64_t>::max());
    uint64_t wb = next<uint64_t>(1, std::numeric_limits<uint64_t>::max());
    bool a_on_positive = ccw(C, A, B) < 0;
 
    // Classify interior points into two sides of the wedge A-C-B for partition.
    // Collinear points on AB are tie-broken along the segment.
    i128 proj_sum = proj_on_ab(A, A, B) + proj_on_ab(B, A, B);
    auto is_positive = [&](const point<long long> &P) -> bool {
        i128 s = wa * ccw(C, A, P) + wb * ccw(C, B, P);
        // Weighted wedge side of P w.r.t. C, A, B.
        if (s != 0)
            return s > 0;
        // P is on line AB: split by projection past the midpoint.
        return 2 * proj_on_ab(P, A, B) > proj_sum;
    };
 
    // Holds C at points[right-2] while classifying interior points in
    // [left+1, right-3].
    if (ci != right - 2)
        std::swap(points[ci], points[right - 2]);
 
    int i = left + 1;
    int j = right - 3;
    while (i < j) {
        if (is_positive(points[i]) == a_on_positive)
            ++i;
        else if (is_positive(points[j]) != a_on_positive)
            --j;
        else {
            std::swap(points[i], points[j]);
            ++i;
            --j;
        }
    }
 
    // After partition:
    // points[left]=A | (A,C)... | C | (C,B)... | points[right-1]=B.
 
    // After the swap, p is the index of C (pivot between the two subpaths).
    int p = i;
    if (i == j and is_positive(points[i]) == a_on_positive)
        ++p;
    std::swap(points[p], points[right - 2]);
 
    // Path A -> C.
    conquer(points, left, p + 1);
    // Path C -> B.
    conquer(points, p, right);
}
 
// Samples k sorted distinct integers from [left, right] uniformly.
// Optimized for performance (pool partial Fisher–Yates or complement path for
// modest ranges; sparse-map fallback otherwise).
// O(k log k); O(right - left) memory when the range is modest.
inline std::vector<long long>
sample_sorted_distinct_in_range(int k, long long left, long long right) {
    long long universe = right - left + 1;
    std::vector<long long> res;
    res.reserve(k);
    if (k == 0)
        return res;
 
    constexpr long long pool_threshold = 8'000'000;
    constexpr long long pool_always_below = 500'000;
 
    if (universe <= pool_threshold and
        (universe <= pool_always_below or k >= universe / 4)) {
        size_t u = universe;
        size_t ks = k;
        std::vector<long long> pool(u);
        std::iota(pool.begin(), pool.end(), left);
        size_t m = ks <= u / 2 ? ks : u - ks;
        for (size_t i = 0; i < m; ++i) {
            size_t j = next<size_t>(i, u - 1);
            std::swap(pool[i], pool[j]);
        }
        if (ks <= u / 2) {
            res.assign(pool.begin(), pool.begin() + ks);
            std::sort(res.begin(), res.end());
        } else {
            std::vector<char> excluded(u, 0);
            for (size_t i = 0; i < m; ++i)
                excluded[pool[i] - left] = 1;
            for (long long v = left; v <= right; ++v)
                if (!excluded[v - left])
                    res.push_back(v);
        }
    } else {
        std::unordered_map<long long, long long> virtual_list;
        virtual_list.reserve(k * 2);
        for (long long i = 0; i < k; ++i) {
            long long j = next<long long>(i, universe - 1);
            long long vi = virtual_list.count(i) ? virtual_list[i] : i;
            long long vj = virtual_list.count(j) ? virtual_list[j] : j;
            virtual_list[j] = vi;
            virtual_list[i] = vj;
            res.push_back(virtual_list[i] + left);
        }
        std::sort(res.begin(), res.end());
    }
    return res;
}
 
// Valtr-style signed edge components along one axis from n sorted distinct
// coordinates. The n differences sum to zero.
inline std::vector<long long>
valtr_edge_components(const std::vector<long long> &sorted_coords) {
    int n = sorted_coords.size();
    std::vector<long long> left, right;
    left.reserve(n / 2);
    right.reserve(n / 2);
    for (int i = 1; i + 1 < n; ++i) {
        if (next(2) == 0)
            left.push_back(sorted_coords[i]);
        else
            right.push_back(sorted_coords[i]);
    }
    long long lo = sorted_coords.front(), hi = sorted_coords.back();
    std::vector<long long> seq;
    seq.reserve(n + 1);
    seq.push_back(lo);
    for (long long v : left)
        seq.push_back(v);
    seq.push_back(hi);
    for (auto it = right.rbegin(); it != right.rend(); ++it)
        seq.push_back(*it);
    seq.push_back(lo);
    std::vector<long long> comps(n);
    for (int i = 0; i < n; ++i)
        comps[i] = seq[i + 1] - seq[i];
    return comps;
}
 
// Drops boundary vertices that are collinear with their cyclic neighbors.
// O(m), m = |points|.
inline std::vector<point<long long>>
simplify_strict_boundary(std::vector<point<long long>> points) {
    int n = points.size();
    if (n < 3)
        return points;
 
    std::vector<point<long long>> strict_points;
    strict_points.reserve(n);
    for (int i = 0; i < n; ++i) {
        if (ccw(points[(i + n - 1) % n], points[i], points[(i + 1) % n]) != 0)
            strict_points.push_back(points[i]);
    }
    return strict_points;
}
 
// Picks k evenly spaced vertices along a longer cyclic boundary.
// O(k).
inline std::vector<point<long long>>
subsample_boundary(const std::vector<point<long long>> &points, int k) {
    int n = points.size();
    if (n <= k)
        return points;
 
    std::vector<point<long long>> sampled_points;
    sampled_points.reserve(k);
    for (int i = 0; i < k; ++i)
        sampled_points.push_back(points[(static_cast<i128>(i) * n) / k]);
    return sampled_points;
}
 
// Random translation so the polygon lies in the box.
// O(|points|).
inline void place_inside_box(std::vector<point<long long>> &points,
                             long long min_coord, long long max_coord) {
    long long width = max_coord - min_coord + 1;
 
    i128 min_x = points[0].x(), max_x = points[0].x();
    i128 min_y = points[0].y(), max_y = points[0].y();
    for (const point<long long> &p : points) {
        min_x = std::min(min_x, static_cast<i128>(p.x()));
        max_x = std::max(max_x, static_cast<i128>(p.x()));
        min_y = std::min(min_y, static_cast<i128>(p.y()));
        max_y = std::max(max_y, static_cast<i128>(p.y()));
    }
 
    i128 span_x = max_x - min_x;
    i128 span_y = max_y - min_y;
    // Random slack keeps the polygon inside the box without filling it.
    i128 shift_x =
        min_coord - min_x +
        next<long long>(0, width - 1 - static_cast<long long>(span_x));
    i128 shift_y =
        min_coord - min_y +
        next<long long>(0, width - 1 - static_cast<long long>(span_y));
 
    for (point<long long> &p : points)
        p = point<long long>(p.x() + shift_x, p.y() + shift_y);
}
 
// Random cyclic shift.
// O(|points|).
inline void randomize_cyclic_shift(std::vector<point<long long>> &points) {
    int rot = next(points.size());
    if (rot > 0)
        std::rotate(points.begin(), points.begin() + rot, points.end());
}
 
// Valtr walk for m edges; bbox minimum translated to the origin.
// O(m log m).
inline std::vector<point<long long>>
valtr_vertices(int m, const std::vector<long long> &x_comp,
               std::vector<long long> y_comp) {
    shuffle(y_comp.begin(), y_comp.end());
 
    std::vector<point<long long>> edges(m);
    // Upper half-plane (positive y, or y = 0 and x > 0) sorts before lower.
    auto upper = [](const point<long long> &p) {
        return p.y() > 0 or (p.y() == 0 and p.x() > 0);
    };
    for (int i = 0; i < m; ++i)
        edges[i] = point<long long>(x_comp[i], y_comp[i]);
 
    std::sort(edges.begin(), edges.end(),
              [&upper](const point<long long> &a, const point<long long> &b) {
                  bool au = upper(a), bu = upper(b);
                  if (au != bu)
                      return au;
                  auto cross = a ^ b;
                  if (cross != 0)
                      return cross > 0;
                  return (a * a) < (b * b);
              });
 
    // Prefix-sum the sorted edge vectors to obtain vertex coordinates.
    i128 cur_x = 0, cur_y = 0;
    std::vector<i128> px(m), py(m);
    for (int i = 0; i < m; ++i) {
        px[i] = cur_x;
        py[i] = cur_y;
        cur_x += edges[i].x();
        cur_y += edges[i].y();
    }
    tgen::detail::tgen_ensure_against_bug(
        cur_x == 0 and cur_y == 0,
        "geometry: random_convex_polygon: walk did not close");
 
    i128 min_x = px[0], min_y = py[0];
    for (int i = 1; i < m; ++i) {
        min_x = std::min(min_x, px[i]);
        min_y = std::min(min_y, py[i]);
    }
 
    // Shift so the bbox minimum is at the origin.
    std::vector<point<long long>> points;
    points.reserve(m);
    for (int i = 0; i < m; ++i)
        points.emplace_back(px[i] - min_x, py[i] - min_y);
    return points;
}
 
} // namespace detail
 
// Generates n vertices of a convex integer polygon inside a box.
// If strict is true, boundary vertices are guaranteed non-collinear when
// generation succeeds; retry count depends on n and width.
// Always returns points in counterclockwise order.
// O(n log n).
inline std::vector<point<long long>>
random_convex_polygon(int n, long long min_coord, long long max_coord,
                      bool strict = false) {
    tgen_ensure(n >= 3,
                "geometry: random_convex_polygon: n must be at least 3");
    tgen_ensure(max_coord >= min_coord,
                "geometry: random_convex_polygon: min_coord must be at most "
                "max_coord");
    tgen_ensure(static_cast<detail::i128>(max_coord) - min_coord + 1 <=
                    std::numeric_limits<long long>::max(),
                "geometry: random_convex_polygon: coordinate range too large");
    long long width = max_coord - min_coord + 1;
    tgen_ensure(
        width >= n,
        "geometry: random_convex_polygon: coordinate range too small for n");
 
    // Valtr walk size: n in weak mode; strict mode uses a larger grid so
    // collinear removal still leaves at least n vertices to subsample.
    int num_coords = n;
    if (strict) {
        // Extra grid lines beyond n: at least 100 (small-n headroom), about
        // n/1000 for large n, and never more than width - n.
        int extra = width <= n ? 0
                               : std::min<long long>(std::max(100, n / 1000),
                                                     width - n);
        num_coords = n + extra;
    }
 
    // Strict mode retries coordinate sampling when simplification leaves < n
    // vertices; weak mode has no failure path, so one attempt always suffices.
    const int max_attempts = strict ? 32 : 1;
    for (int i = 0; i < max_attempts; ++i) {
        // Build a convex lattice polygon on [0, width - 1]^2, then translate.
        std::vector<long long> x_sorted =
            detail::sample_sorted_distinct_in_range(num_coords, 0, width - 1);
        std::vector<long long> y_sorted =
            detail::sample_sorted_distinct_in_range(num_coords, 0, width - 1);
        std::vector<long long> x_comp = detail::valtr_edge_components(x_sorted);
        std::vector<long long> y_comp = detail::valtr_edge_components(y_sorted);
 
        std::vector<point<long long>> points =
            detail::valtr_vertices(num_coords, x_comp, std::move(y_comp));
 
        if (strict) {
            std::vector<point<long long>> simplified =
                detail::simplify_strict_boundary(std::move(points));
            // Tight boxes can leave too few vertices -> resample coordinates.
            if (static_cast<int>(simplified.size()) < n)
                continue;
 
            points = detail::subsample_boundary(simplified, n);
        }
 
        detail::place_inside_box(points, min_coord, max_coord);
        detail::randomize_cyclic_shift(points);
        return points;
    }
 
    // Generation failed.
    throw tgen::detail::error(
        "geometry: random_convex_polygon: generation failed: coordinate "
        "range too small for n");
}
 
// Random simple polygon through given distinct points.
// Collinear triples are allowed; fails if all points are collinear.
// Always returns vertices in counterclockwise order.
// O(n log n) expected if points are "random", O(n^2) worst case.
inline std::vector<point<long long>> random_simple_polygon_through_points(
    const std::vector<point<long long>> &points) {
    int n = points.size();
    tgen_ensure(n >= 3,
                "geometry: random_simple_polygon_through_points: need at "
                "least 3 points");
    tgen_ensure(
        static_cast<int>(
            std::set<point<long long>>(points.begin(), points.end()).size()) ==
            n,
        "geometry: random_simple_polygon_through_points: points must "
        "be distinct");
 
    int idx_a = 0, idx_b = 0;
    for (int i = 1; i < n; ++i) {
        if (points[i] < points[idx_a])
            idx_a = i;
        if (points[idx_b] < points[i])
            idx_b = i;
    }
    point<long long> A = points[idx_a], B = points[idx_b];
 
    bool all_collinear = true;
    for (int i = 0; i < n; ++i) {
        if (i == idx_a or i == idx_b)
            continue;
        if (detail::ccw(A, B, points[i]) != 0) {
            all_collinear = false;
            break;
        }
    }
    tgen_ensure(!all_collinear,
                "geometry: random_simple_polygon_through_points: all points "
                "are collinear; no simple polygon exists");
 
    // Keep points collinear with AB on the chain that has no other points on
    // its side, so AB is split through those vertices instead of crossing them
    // later.
    int negative_count = 0;
    for (int i = 0; i < n; ++i) {
        if (i == idx_a or i == idx_b)
            continue;
        if (detail::ccw(A, B, points[i]) < 0)
            ++negative_count;
    }
 
    std::vector<point<long long>> chain;
    chain.push_back(A);
    int left_count = 0;
    for (int i = 0; i < n; ++i) {
        if (i == idx_a or i == idx_b)
            continue;
        detail::i128 side = detail::ccw(A, B, points[i]);
        if (side < 0 or (side == 0 and negative_count == 0)) {
            chain.push_back(points[i]);
            ++left_count;
        }
    }
    chain.push_back(B);
    for (int i = 0; i < n; ++i) {
        if (i == idx_a or i == idx_b)
            continue;
        detail::i128 side = detail::ccw(A, B, points[i]);
        if (side > 0 or (side == 0 and negative_count != 0))
            chain.push_back(points[i]);
    }
    chain.push_back(A);
 
    int n1 = 2 + left_count;
    // Upper chain: A -> B.
    detail::conquer(chain, 0, n1);
    // Lower chain: B -> A.
    detail::conquer(chain, n1 - 1, chain.size());
 
    // Cyclic vertex order: chain[1..n1) then chain[n1..end) (skip each path's
    // start vertex).
    std::vector<point<long long>> poly;
    poly.insert(poly.end(), chain.begin() + 1, chain.begin() + n1);
    poly.insert(poly.end(), chain.begin() + n1, chain.end());
    return poly;
}
 
namespace detail {
 
// Samples n distinct integer points in [min_coord, max_coord]^2.
// O(n log n).
inline std::vector<point<long long>>
random_distinct_points_in_box(int n, long long min_coord, long long max_coord) {
    long long width = max_coord - min_coord;
    i128 side_128 = width + 1;
    i128 universe = side_128 * side_128;
    tgen_ensure(universe <= std::numeric_limits<long long>::max(),
                "geometry: random_simple_polygon: coordinate range too large");
    long long side = side_128;
    tgen_ensure(universe >= n,
                "geometry: random_simple_polygon: coordinate range too small "
                "for n distinct points");
 
    // Decodes a linear grid key (x * side + y) to a point.
    auto decode = [&](long long key) -> point<long long> {
        return point<long long>(min_coord + key / side, min_coord + key % side);
    };
 
    // Repeats until all generated points are not collinear.
    // Runs O(1) expected times.
    while (true) {
        std::vector<long long> keys =
            distinct_range<long long>(0, universe - 1).gen_list(n).to_std();
 
        std::vector<point<long long>> points;
        points.reserve(n);
        for (long long key : keys)
            points.push_back(decode(key));
 
        // Checks if the points are not all collinear.
        for (int i = 2; i < n; ++i) {
            if (ccw(points[0], points[1], points[i]) != 0)
                return points;
        }
    }
}
 
// Axis-aligned edge data for a CCW polygon (interior on the left).
// O(1).
struct ortho_poly_edge {
    long long len;
    bool horiz;
    long long fixed;
    long long lo, hi;
    int out_x, out_y;
};
 
// True if a, b, c lie on one horizontal or vertical line.
// O(1).
inline bool ortho_axis_collinear(const point<long long> &a,
                                 const point<long long> &b,
                                 const point<long long> &c) {
    return (a.x() == b.x() and b.x() == c.x()) or
           (a.y() == b.y() and b.y() == c.y());
}
 
// Edge i of poly: orientation, span, exterior normal.
// O(1).
inline ortho_poly_edge
ortho_analyze_edge(const std::vector<point<long long>> &poly, int i) {
    int m = poly.size();
    point<long long> a = poly[i], b = poly[(i + 1) % m];
    ortho_poly_edge e{};
    if (a.y() == b.y()) {
        e.horiz = true;
        e.fixed = a.y();
        e.lo = std::min(a.x(), b.x());
        e.hi = std::max(a.x(), b.x());
        e.out_x = 0;
        e.out_y = a.x() < b.x() ? -1 : 1;
    } else {
        e.fixed = a.x();
        e.lo = std::min(a.y(), b.y());
        e.hi = std::max(a.y(), b.y());
        e.out_x = a.y() < b.y() ? 1 : -1;
    }
    e.len = e.hi - e.lo;
    return e;
}
 
// True if open axis-aligned segments (a, b) and (c, d) properly cross or
// overlap (excluding shared endpoints).
// O(1).
inline bool ortho_open_seg_cross(const point<long long> &a,
                                 const point<long long> &b,
                                 const point<long long> &c,
                                 const point<long long> &d) {
    if (a.y() == b.y() and c.y() == d.y()) {
        if (a.y() != c.y())
            return false;
        long long lo1 = std::min(a.x(), b.x()), hi1 = std::max(a.x(), b.x());
        long long lo2 = std::min(c.x(), d.x()), hi2 = std::max(c.x(), d.x());
        return lo1 < hi2 and lo2 < hi1;
    }
    if (a.x() == b.x() and c.x() == d.x()) {
        if (a.x() != c.x())
            return false;
        long long lo1 = std::min(a.y(), b.y()), hi1 = std::max(a.y(), b.y());
        long long lo2 = std::min(c.y(), d.y()), hi2 = std::max(c.y(), d.y());
        return lo1 < hi2 and lo2 < hi1;
    }
    if (a.y() == b.y() and c.x() == d.x()) {
        long long hx = a.y(), vx = c.x();
        long long hlo = std::min(a.x(), b.x()), hhi = std::max(a.x(), b.x());
        long long vlo = std::min(c.y(), d.y()), vhi = std::max(c.y(), d.y());
        return hlo < vx and vx < hhi and vlo < hx and hx < vhi;
    }
    if (a.x() == b.x() and c.y() == d.y()) {
        long long vx = a.x(), hy = c.y();
        long long vlo = std::min(a.y(), b.y()), vhi = std::max(a.y(), b.y());
        long long hlo = std::min(c.x(), d.x()), hhi = std::max(c.x(), d.x());
        return vlo < hy and hy < vhi and hlo < vx and vx < hhi;
    }
    return false;
}
 
// Integral ray-crossing point-in-polygon.
// O(|poly|).
inline bool ortho_point_inside(const std::vector<point<long long>> &poly,
                               point<long long> p) {
    int m = poly.size();
    bool inside = false;
    for (int i = 0, j = m - 1; i < m; j = i++) {
        point<long long> a = poly[i], b = poly[j];
        if ((a.y() > p.y()) != (b.y() > p.y())) {
            i128 x_cross = i128(b.x() - a.x()) * (p.y() - a.y()) -
                           i128(p.x() - a.x()) * (b.y() - a.y());
            if ((a.y() < b.y()) ? x_cross > 0 : x_cross < 0)
                inside = !inside;
        }
    }
    return inside;
}
 
// True if p lies strictly in the open segment (a, b).
// O(1).
inline bool ortho_point_strictly_interior(point<long long> p,
                                          point<long long> a,
                                          point<long long> b) {
    if (a.y() == b.y()) {
        if (p.y() != a.y())
            return false;
        long long lo = std::min(a.x(), b.x()), hi = std::max(a.x(), b.x());
        return lo < p.x() and p.x() < hi;
    }
    if (a.x() == b.x()) {
        if (p.x() != a.x())
            return false;
        long long lo = std::min(a.y(), b.y()), hi = std::max(a.y(), b.y());
        return lo < p.y() and p.y() < hi;
    }
    return false;
}
 
// True if p lies on the closed segment [a, b].
// O(1).
inline bool ortho_point_on_segment(point<long long> p, point<long long> a,
                                   point<long long> b) {
    return p == a or p == b or ortho_point_strictly_interior(p, a, b);
}
 
// True if splicing add between A and B on edge_i is valid: no vertex of add
// coincides with poly, no boundary self-contact (forward or reverse
// T-junctions, collinear overlaps), new segments do not cross other boundary
// edges, and inward notches keep all of add inside poly.
// O(|poly|), assuming |add| is O(1).
inline bool ortho_bump_valid(const std::vector<point<long long>> &poly,
                             point<long long> A, point<long long> B,
                             const std::vector<point<long long>> &add,
                             int edge_i, bool inward) {
    int m = poly.size();
 
    for (point<long long> v : add)
        for (point<long long> q : poly)
            if (v == q)
                return false;
 
    // Forward T-junction: new vertex on a non-incident edge interior, or on
    // edge_i but outside the replaced subsegment [A, B].
    for (point<long long> v : add) {
        for (int j = 0; j < m; ++j) {
            point<long long> c = poly[j], d = poly[(j + 1) % m];
            if (j == edge_i) {
                if (ortho_point_on_segment(v, c, d) and
                    !ortho_point_on_segment(v, A, B))
                    return false;
            } else if (ortho_point_strictly_interior(v, c, d)) {
                return false;
            }
        }
    }
 
    auto seg_ok = [&](point<long long> s0, point<long long> s1) {
        for (int j = 0; j < m; ++j) {
            if (j == edge_i)
                continue;
            point<long long> c = poly[j], d = poly[(j + 1) % m];
            if (ortho_open_seg_cross(s0, s1, c, d))
                return false;
        }
        for (int k = 0; k < m; ++k) {
            point<long long> q = poly[k];
            if (q == s0 or q == s1 or q == A or q == B)
                continue;
            if (ortho_point_strictly_interior(q, s0, s1))
                return false;
        }
        return true;
    };
 
    point<long long> prev = A;
    for (point<long long> v : add) {
        if (!seg_ok(prev, v))
            return false;
        prev = v;
    }
    if (!seg_ok(prev, B))
        return false;
 
    if (inward) {
        for (point<long long> v : add)
            if (!ortho_point_inside(poly, v))
                return false;
    }
 
    return true;
}
 
// Splices a rectangular tab or notch on edge edge_i over [lo, hi], extending
// depth units perpendicular to the edge.
// O(|poly|).
inline bool ortho_bump_edge(std::vector<point<long long>> &poly, int edge_i,
                            const ortho_poly_edge &e, long long lo,
                            long long hi, long long depth, bool inward) {
    int m = poly.size();
    point<long long> A = poly[edge_i], B = poly[(edge_i + 1) % m];
 
    int step_x = inward ? -e.out_x : e.out_x;
    int step_y = inward ? -e.out_y : e.out_y;
 
    std::vector<point<long long>> add;
    if (e.horiz) {
        long long y = e.fixed, y2 = y + step_y * depth;
        if (A.x() < B.x()) {
            if (lo > A.x())
                add.emplace_back(lo, y);
            add.emplace_back(lo, y2);
            add.emplace_back(hi, y2);
            if (hi < B.x())
                add.emplace_back(hi, y);
        } else {
            if (hi < A.x())
                add.emplace_back(hi, y);
            add.emplace_back(hi, y2);
            add.emplace_back(lo, y2);
            if (lo > B.x())
                add.emplace_back(lo, y);
        }
    } else {
        long long x = e.fixed, x2 = x + step_x * depth;
        if (A.y() < B.y()) {
            if (lo > A.y())
                add.emplace_back(x, lo);
            add.emplace_back(x2, lo);
            add.emplace_back(x2, hi);
            if (hi < B.y())
                add.emplace_back(x, hi);
        } else {
            if (hi < A.y())
                add.emplace_back(x, hi);
            add.emplace_back(x2, hi);
            add.emplace_back(x2, lo);
            if (lo > B.y())
                add.emplace_back(x, lo);
        }
    }
    if (!ortho_bump_valid(poly, A, B, add, edge_i, inward))
        return false;
 
    poly.insert(poly.begin() + edge_i + 1, add.begin(), add.end());
    return true;
}
 
// Picks edge i with probability proportional to
// e.len * (4 + min(global_timestamp - last_used[i], 8)).
// O(|poly|).
inline int ortho_pick_poly_edge(const std::vector<point<long long>> &poly,
                                std::vector<int> &last_used, int &time_stamp) {
    int m = poly.size();
    if (last_used.size() != static_cast<size_t>(m)) {
        last_used.assign(m, 0);
        time_stamp = 0;
    }
    std::vector<long long> weights(m);
    long long total = 0;
    for (int i = 0; i < m; ++i) {
        ortho_poly_edge e = ortho_analyze_edge(poly, i);
        weights[i] = e.len * (4 + std::min(time_stamp - last_used[i], 8));
        total += weights[i];
    }
    long long pick = next<long long>(0, total - 1);
    for (int i = 0; i < m; ++i) {
        pick -= weights[i];
        if (pick < 0) {
            last_used[i] = ++time_stamp;
            return i;
        }
    }
    last_used[m - 1] = ++time_stamp;
    return m - 1;
}
 
// One random inflate/cut attempt.
// O(|poly|).
inline bool ortho_try_bump(std::vector<point<long long>> &poly, int n,
                           std::vector<int> &last_used, int &time_stamp,
                           bool outward_only = false) {
    if (poly.size() < 3)
        return false;
 
    int ei = ortho_pick_poly_edge(poly, last_used, time_stamp);
    ortho_poly_edge e = ortho_analyze_edge(poly, ei);
 
    // Edge subdivision can leave length-1 segments; a tab needs span >= 2.
    if (e.len < 2)
        return false;
 
    // Random subinterval [lo, lo + span] on the edge, with 2 <= span <= e.len.
    long long span = next<long long>(2, e.len);
    long long lo = next<long long>(e.lo, e.hi - span);
 
    // max_depth: cap on perpendicular tab/notch height (~sqrt(n), in [2, 12]).
    // depth: actual height; shallow usually, up to max_depth 10% of the time.
    long long max_depth =
        std::clamp<long long>(std::sqrt(n) / 2 + 2, 2LL, 12LL);
    long long depth =
        next(10) == 0 ? next<long long>(std::max(2LL, max_depth / 2), max_depth)
                      : next<long long>(1, std::max(2LL, max_depth / 3));
 
    bool inward = !outward_only and next(4) == 0;
    return ortho_bump_edge(poly, ei, e, lo, lo + span, depth, inward);
}
 
// Drops axis-aligned collinear vertices.
// O(n).
inline std::vector<point<long long>>
ortho_simplify_collinear(std::vector<point<long long>> poly) {
    int n = poly.size();
    if (n < 3)
        return poly;
    std::vector<point<long long>> out;
    out.reserve(n);
    for (int i = 0; i < n; ++i) {
        if (!ortho_axis_collinear(poly[(i + n - 1) % n], poly[i],
                                  poly[(i + 1) % n]))
            out.push_back(poly[i]);
    }
    return out.size() >= 3 ? out : poly;
}
 
// Removes one collinear vertex.
// O(n).
inline bool ortho_remove_one_collinear(std::vector<point<long long>> &poly) {
    int n = poly.size();
    if (n < 4)
        return false;
    for (int i = 0; i < n; ++i) {
        if (!ortho_axis_collinear(poly[(i + n - 1) % n], poly[i],
                                  poly[(i + 1) % n]))
            continue;
        poly.erase(poly.begin() + i);
        return true;
    }
    return false;
}
 
// Inserts collinear vertices on straight edges until size reaches `target` (or
// no edge has spare integer points). Preserves boundary order.
// O(target).
inline void ortho_fill_collinear(std::vector<point<long long>> &poly,
                                 int target) {
    int need = target - poly.size();
    if (need <= 0)
        return;
 
    std::vector<point<long long>> out;
    int m = poly.size();
    out.reserve(poly.size() + need);
 
    for (int i = 0; i < m; ++i) {
        point<long long> a = poly[i], b = poly[(i + 1) % m];
        out.push_back(a);
        if (need <= 0)
            continue;
 
        ortho_poly_edge e = ortho_analyze_edge(poly, i);
        long long cap = e.len - 1;
        if (cap <= 0)
            continue;
 
        // Adds `take` collinear vertices to the boundary.
        long long take = std::min<long long>(need, cap);
        bool forward = e.horiz ? a.x() < b.x() : a.y() < b.y();
        for (long long k = 0; k < take; ++k) {
            long long off = (k + 1) * (cap + 1) / (take + 1);
            long long coord = forward ? e.lo + off : e.hi - off;
            if (e.horiz)
                out.push_back({coord, e.fixed});
            else
                out.push_back({e.fixed, coord});
        }
        need -= take;
    }
    poly.swap(out);
}
 
// Inserts outward depth-1 corrugation tabs on straight edges until size reaches
// `target` (or no edge has room for another tab). Preserves boundary order.
// Assumes the scale-up step left every edge on a grid with spacing >= 4.
// Each tab perturbs a coordinate by 1 and keeps a >= 2 margin from both
// corners, so tabs can only meet tabs on the same or a directly facing edge.
// The result is therefore simple and free of collinear triples by construction.
// O(target).
inline void ortho_fill_corrugation(std::vector<point<long long>> &poly,
                                   size_t target) {
    size_t n_sz = poly.size();
    if (n_sz >= target)
        return;
 
    size_t extra_left = target - n_sz;
 
    std::vector<point<long long>> out;
    out.reserve(target);
    int n = poly.size();
 
    for (int i = 0; i < n; ++i) {
        point<long long> a = poly[i], b = poly[(i + 1) % n];
        out.push_back(a);
 
        if (extra_left < 4)
            continue;
 
        ortho_poly_edge e = ortho_analyze_edge(poly, i);
        if (e.len < 5)
            continue;
 
        long long dir = (e.horiz ? a.x() < b.x() : a.y() < b.y()) ? 1 : -1;
        long long start = e.horiz ? a.x() : a.y();
        long long end = e.horiz ? b.x() : b.y();
 
        for (long long pos = start + 2 * dir;
             extra_left >= 4 and (pos - end) * dir <= -3; pos += 2 * dir) {
            if (e.horiz) {
                long long y2 = e.fixed + e.out_y;
                out.push_back({pos, e.fixed});
                out.push_back({pos, y2});
                out.push_back({pos + dir, y2});
                out.push_back({pos + dir, e.fixed});
            } else {
                long long x2 = e.fixed + e.out_x;
                out.push_back({e.fixed, pos});
                out.push_back({x2, pos});
                out.push_back({x2, pos + dir});
                out.push_back({e.fixed, pos + dir});
            }
            extra_left -= 4;
        }
    }
 
    poly.swap(out);
}
 
// CCW square seed plus boundary inflate/cut to about n vertices.
// O(n^2) for n <= 1000; O(n) otherwise.
inline std::vector<point<long long>> build_orthogonal_polygon(int n,
                                                              bool strict) {
    bool scale_up = n > 1000;
 
    long long side = std::max<long long>(3, std::sqrt(n));
    if (scale_up)
        side = std::clamp(static_cast<long long>(2 * std::sqrt(std::sqrt(n))),
                          8LL, 64LL);
    std::vector<point<long long>> poly = {
        {0, 0}, {side, 0}, {side, side}, {0, side}};
 
    int target_ops = scale_up ? std::max(1, static_cast<int>(4 * side - 4) / 2)
                              : std::max(1, (n - 4) / 2);
 
    if (scale_up)
        target_ops = std::min(
            target_ops,
            400 + static_cast<int>(4 * std::sqrt(static_cast<double>(side))));
 
    int failure_limit = std::min(target_ops * 8, 2000);
 
    std::vector<int> last_used;
    int time_stamp = 0, consecutive_failures = 0;
    for (int ops = 0; ops < target_ops;) {
        if (!scale_up and poly.size() + 2 > static_cast<size_t>(n))
            break;
 
        if (ortho_try_bump(poly, n, last_used, time_stamp, scale_up)) {
            ++ops;
            consecutive_failures = 0;
        } else if (++consecutive_failures >= failure_limit) {
            break;
        }
    }
    if (strict)
        poly = ortho_simplify_collinear(std::move(poly));
 
    if (scale_up) {
        while (poly.size() > static_cast<size_t>(n) and
               ortho_remove_one_collinear(poly))
            ;
        long long upscale = std::max(4LL, ((n + 3) / 4 + side - 1) / side);
        for (point<long long> &p : poly)
            p = {p.x() * upscale, p.y() * upscale};
        if (strict)
            ortho_fill_corrugation(poly, n);
        else
            ortho_fill_collinear(poly, n);
    } else if (!strict)
        ortho_fill_collinear(poly, n);
 
    return poly;
}
 
} // namespace detail
 
// Random simple polygon.
// If strict, vertex set has no three collinear points
// (random_points_general_position); otherwise samples distinct grid points
// (collinear triples allowed), so it might be the case that (polygon[i],
// polygon[i+1], and polygon[i+2]) are collinear. Polygonizes via
// random_simple_polygon_through_points. Always counterclockwise.
// O(n log n) expected.
inline std::vector<point<long long>>
random_simple_polygon(int n, long long min_coord, long long max_coord,
                      bool strict = false) {
    tgen_ensure(n >= 3,
                "geometry: random_simple_polygon: n must be at least 3");
    tgen_ensure(max_coord >= min_coord,
                "geometry: random_simple_polygon: min_coord must be at most "
                "max_coord");
    tgen_ensure(static_cast<detail::i128>(max_coord) - min_coord <=
                    std::numeric_limits<long long>::max(),
                "geometry: random_simple_polygon: coordinate range too large");
 
    std::vector<point<long long>> points =
        strict ? random_points_general_position(n, min_coord, max_coord)
               : detail::random_distinct_points_in_box(n, min_coord, max_coord);
    return random_simple_polygon_through_points(points);
}
 
// Random orthogonal simple polygon, CCW. Each local bump/scale/fill step
// preserves full simplicity, so the result is valid by construction.
// Exactly n vertices when !strict; at most n when strict (near n for n > 1000).
// O(n^2) for n <= 1000; O(n) otherwise.
inline std::vector<point<long long>>
random_orthogonal_polygon(int n, long long min_coord, long long max_coord,
                          bool strict = false) {
    tgen_ensure(n >= 4,
                "geometry: random_orthogonal_polygon: n must be at least 4");
    tgen_ensure(max_coord >= min_coord,
                "geometry: random_orthogonal_polygon: min_coord must be at "
                "most max_coord");
    tgen_ensure(static_cast<detail::i128>(max_coord) - min_coord + 1 <=
                    std::numeric_limits<long long>::max(),
                "geometry: random_orthogonal_polygon: coordinate range too "
                "large");
    long long width = max_coord - min_coord + 1;
    tgen_ensure(width >= 4,
                "geometry: random_orthogonal_polygon: coordinate range too "
                "small");
 
    long long min_side = std::max<long long>(3, std::sqrt(n));
    if (n > 1000)
        min_side = std::max<long long>(min_side, (n + 3) / 4);
    tgen_ensure(min_side < width,
                "geometry: random_orthogonal_polygon: coordinate range too "
                "small");
 
    for (int attempt = 0; attempt < 8; ++attempt) {
        std::vector<point<long long>> poly =
            detail::build_orthogonal_polygon(n, strict);
 
        if (!strict and poly.size() != static_cast<size_t>(n))
            continue;
 
        detail::i128 min_x = poly[0].x(), max_x = poly[0].x();
        detail::i128 min_y = poly[0].y(), max_y = poly[0].y();
        for (point<long long> p : poly) {
            min_x = std::min(min_x, detail::i128(p.x()));
            max_x = std::max(max_x, detail::i128(p.x()));
            min_y = std::min(min_y, detail::i128(p.y()));
            max_y = std::max(max_y, detail::i128(p.y()));
        }
        if (max_x - min_x >= width or max_y - min_y >= width)
            continue;
 
        detail::place_inside_box(poly, min_coord, max_coord);
        detail::randomize_cyclic_shift(poly);
        return poly;
    }
 
    throw tgen::detail::error(
        "geometry: random_orthogonal_polygon: generation failed");
}
 
} // namespace geometry
 
/************
 *          *
 *   HACK   *
 *          *
 ************/
 
namespace hack {
 
namespace detail {
 
using namespace tgen::detail;
 
// Computes polynomial hash of a string.
// O(|s|).
inline int hash_string(const std::string &s, int base, int mod) {
    long long h = 0;
    for (char c : s)
        h = (h * base + c - 'a' + 1) % mod;
    return h;
}
 
// Estimates the length of the string to very likely have a collision.
inline int estimate_length(int alphabet_size, int mod) {
    // Magic constants.
    double base_len = 2.5 * std::log(std::sqrt(mod));
    double scale = std::log(alphabet_size) / std::log(2.0);
    double adjusted = base_len / std::max(1.0, scale * 0.7);
 
    return static_cast<int>(std::ceil(adjusted));
}
 
// Collides two strings to have the same polynomial hash.
// O(sqrt(mod) log(mod)) with high probability.
inline std::pair<std::string, std::string>
birthday_attack(const std::vector<std::string> &alphabet, int base, int mod) {
    tgen_ensure(0 < base and base < mod,
                "birthday_attack: base must be in (0, mod)");
    std::map<uint64_t, std::vector<int>> seen;
    int length = estimate_length(alphabet.size(), mod);
 
    while (true) {
        std::vector<int> seq(length);
 
        std::string s;
 
        for (int i = 0; i < length; ++i) {
            seq[i] = next<int>(0, alphabet.size() - 1);
            s += alphabet[seq[i]];
        }
 
        int h = hash_string(s, base, mod);
 
        auto it = seen.find(h);
        if (it != seen.end() and it->second != seq) {
            std::string a, b;
 
            for (int x : it->second)
                a += alphabet[x];
            for (int x : seq)
                b += alphabet[x];
 
            if (a != b)
                return {a, b};
        }
 
        seen[h] = seq;
    }
}
 
// Tried to find correct multipliers for unordered_map/set to force
// collisions. O(1).
inline std::set<long long> std_hash_multipliers() {
    std::set<long long> multipliers = {85229};
 
    // Codeforces GCC GNU G++17 7.3.0 case.
    bool codeforces_gcc_case = true;
    if (cpp.version_ != 0 and cpp.version_ != 17)
        codeforces_gcc_case = false;
    if (compiler.kind_ != compiler_kind::unknown and
        compiler.kind_ != compiler_kind::gcc)
        codeforces_gcc_case = false;
    if (compiler.major_ > 7)
        codeforces_gcc_case = false;
 
    if (codeforces_gcc_case)
        multipliers.insert(107897);
 
    return multipliers;
}
 
} // namespace detail
 
// Fetches prefix of length n of the string "abacabadabacabae...".
// O(n).
inline std::string abacaba(int n) {
    tgen_ensure(n > 0, "str: size must be positive");
    std::string str = "a";
    char c = 'a';
    while (static_cast<int>(str.size()) < n) {
        int prev_size = str.size();
        str += ++c;
        for (int j = 0; j < prev_size and static_cast<int>(str.size()) < n; ++j)
            str += str[j];
    }
    return str;
}
 
// Two strings that have same polynomial hash for any base, for
// mod = power of 2 up to 2^64.
// Thue–Morse.
// O(1).
inline std::pair<std::string, std::string> unsigned_polynomial_hash() {
    std::string a, b;
    int size = 1 << 10;
    for (int i = 0; i < size; ++i) {
        a += 'a' + math::detail::popcount(i) % 2;
        b += 'a' + ('b' - a[i]);
    }
    return {a, b};
}
 
// Collides two strings to have the same polynomial hash.
// O(sqrt(mod) log(mod)) with high probability.
// 0 < base < mod.
inline std::pair<std::string, std::string> polynomial_hash(int alphabet_size,
                                                           int base, int mod) {
    tgen_ensure(alphabet_size > 1,
                "hack: polynomial_hash: alphabet size must be greater "
                "than 1");
    tgen_ensure(0 < base and base < mod,
                "hack: polynomial_hash: base must be in (0, mod)");
 
    std::vector<std::string> alphabet(alphabet_size);
    for (int i = 0; i < alphabet_size; ++i)
        alphabet[i] = std::string(1, 'a' + i);
    std::iota(alphabet.begin(), alphabet.end(), 'a');
    return detail::birthday_attack(alphabet, base, mod);
}
 
// Collides two strings to have the same polynomial hash for multiple bases
// and mods (up to 2 pairs).
// O(sqrt(mod) log^2 (mod)) with high probability,
// with mod = max(mod_1, mod_2).
inline std::pair<std::string, std::string>
polynomial_hash(int alphabet_size, std::vector<int> bases,
                std::vector<int> mods) {
    tgen_ensure(bases.size() == mods.size(),
                "hack: polynomial_hash: bases and mods must have the same "
                "size");
    tgen_ensure(bases.size() > 0,
                "hack: polynomial_hash: must have at least one (base, mod) "
                "pair");
    tgen_ensure(bases.size() <= 2,
                "hack: polynomial_hash: multi-hash hack only supported "
                "for up to 2 (base, mod) pairs");
 
    std::vector<std::string> alphabet(alphabet_size);
    for (int i = 0; i < alphabet_size; ++i)
        alphabet[i] = std::string(1, 'a' + i);
    auto [S1, T1] = detail::birthday_attack(alphabet, bases[0], mods[0]);
    if (bases.size() == 1)
        return {S1, T1};
    return detail::birthday_attack({S1, T1}, bases[1], mods[1]);
}
 
// Returns a list of integers for unordered_map/set to force collisions.
// O(size).
inline std::vector<long long> std_unordered(int size) {
    tgen_ensure(size > 0, "hack: std_unordered: size must be positive");
    std::set<long long> multipliers = detail::std_hash_multipliers();
    long long mult = 1;
    std::set<long long>::iterator it = multipliers.begin();
 
    std::vector<long long> list;
    while (static_cast<int>(list.size()) < size) {
        list.push_back(mult * (*it));
        ++it;
        if (it == multipliers.end()) {
            it = multipliers.begin();
            ++mult;
        }
    }
    return list;
}
 
// Returns queries that force \Theta(q sqrt n) asymptotic
// for Mo algorithm for offline range queries.
// Forces \Theta(q sqrt n) pointer moves for any ordering.
// O(n log n + q).
inline std::vector<std::pair<int, int>> mo_worst_case(int n, int q) {
    std::set<std::pair<int, int>> queries;
 
    // Adversarial case.
    int sq = std::sqrt(n);
    for (int i = 0; i < sq; ++i) {
        for (int j = i; j < sq; ++j) {
            if (i * sq < n and j * sq < n)
                queries.emplace(i * sq, j * sq);
        }
    }
 
    // Push extra queries.
    for (int i = 0; i < n; ++i)
        if (queries.size() < size_t(q)) {
            queries.emplace(0, i);
            queries.emplace(i, i);
            queries.emplace(i, n - 1);
        }
 
    std::vector<std::pair<int, int>> pool(queries.begin(), queries.end());
    while (pool.size() < size_t(q)) {
        int l = next(0, n - 1);
        pool.emplace_back(l, next(l, n - 1));
    }
 
    return choose(shuffled(pool), q);
}
 
// Returns list of strings that have a high cost to insert in a std::set.
// Forces cost \Theta(size log(size)).
// Generates: {b, ab, aab, aaab, ...}.
// O(size log(size)).
inline std::vector<std::string> string_set_worst_case(int size) {
    std::vector<std::string> list;
    int k = 0, left = size;
    while (left > 0) {
        int cur_size = std::min(left, k + 1);
        left -= cur_size;
 
        char right_char = cur_size == k + 1 ? 'b' : 'c';
        list.push_back(std::string(cur_size - 1, 'a') + right_char);
 
        ++k;
    }
    return tgen::shuffled(list);
}
 
// Graph for Dijkstra implementations that relax with <= instead of <.
// Unit-weight layered graph: 0 -> {1,2}, then disjoint
// 2x2 gadgets (i,i+1) -> {i+2,i+3} for i = 1,3,5,... Many vertices share the
// same dist from 0; with `d + w <= dist[j]` each pop re-relaxes the whole
// frontier below it. m = 2(n - 2) edges.
// O(n).
inline egraph<int>::value non_strict_relaxation_dijkstra_bug(int n) {
    tgen_ensure(
        n >= 3,
        "hack: non_strict_relaxation_dijkstra_bug: needs at least 3 vertices");
 
    egraph<int>::value g(n, {}, true);
    g.edge_weighted();
    g.add_edge(0, 1, 1);
    g.add_edge(0, 2, 1);
    for (int i = 1; i + 2 < n; i += 2) {
        g.add_edge(i, i + 2, 1);
        if (i + 3 < n)
            g.add_edge(i, i + 3, 1);
 
        g.add_edge(i + 1, i + 2, 1);
        if (i + 3 < n)
            g.add_edge(i + 1, i + 3, 1);
    }
 
    return g.shuffle_except({0});
}
 
// Graph for Dijkstra implementations that do not skip stale heap entries
// (`if (d > dist[i]) continue`).
// Hub mid = n/2: star 0 -> 1..mid-1 (weights 1..mid-1), funnel i -> mid
// (weights 1,3,5,...), then mid -> mid+1.. (weight 1).
// Without a stale-heap check, mid and its in-neighbors are re-popped and
// re-relax.
// m = n + mid - 3 edges, mid = floor(n/2).
// O(n).
inline egraph<int>::value stale_heap_dijkstra_bug(int n) {
    tgen_ensure(n >= 4,
                "hack: stale_heap_dijkstra_bug: needs at least 4 vertices");
 
    int mid = n / 2;
    egraph<int>::value g(n, {}, true);
    g.edge_weighted();
    for (int i = 1; i < mid; ++i)
        g.add_edge(0, i, i);
    for (int i = 1; i < mid; ++i)
        g.add_edge(i, mid, 2 * (mid - i) - 1);
    for (int i = mid + 1; i < n; ++i)
        g.add_edge(mid, i, 1);
 
    return g.shuffle_except({0});
}
 
// Worst-case for FIFO-SPFA.
// Forces Omega(n^2) from vertex 0 (Theta(n*m), m = 2n - 3).
// Upper chain ai -> a(i+1) weight 1; lower chain bi -> b(i+1) weight 0;
// vertical ai -> bi weight 0; cross bi -> a(i+1) weight 1. Upper chain sets
// loose dist first; cross edges from settled bi then improve a(i+1).
// m = 2n - 3.
// O(n).
inline egraph<int>::value spfa(int n) {
    tgen_ensure(n >= 2, "hack: spfa: n must be at least 2");
    tgen_ensure(n % 2 == 0, "hack: spfa: n must be even");
 
    egraph<int>::value g(n, {}, true);
    g.edge_weighted();
 
    const int k = n / 2;
    for (int i = 0; i + 1 < k; ++i)
        g.add_edge(i, i + 1, 1);
    for (int i = 0; i + 1 < k; ++i)
        g.add_edge(k + i, k + i + 1, 0);
    for (int i = 0; i < k; ++i)
        g.add_edge(i, k + i, 0);
    for (int i = 0; i + 1 < k; ++i)
        g.add_edge(k + i, i + 1, 1);
 
    return g.shuffle_except({0});
}
 
// Zadeh (1972) anti-shortest-paths flow network for Edmonds-Karp and Dinitz.
// Source is vertex 0; sink is vertex 4l + 2k + 1.
// n = 4l + 2k + 2, m = 6l + 4k + k^2 - 4.
// O(l + k^2).
inline egraph<int>::value dinitz_worst_case(int k, int l) {
    tgen_ensure(k >= 1, "hack: dinitz_worst_case: k must be at least 1");
    tgen_ensure(l >= 1, "hack: dinitz_worst_case: l must be at least 1");
 
    const int p1 = 2 * l - 1;
    const int p2 = 2 * l;
    const int q1 = 2 * l + 1;
    const int q2 = 2 * l + 2;
    const int n = 4 * l + 2 * k + 2;
 
    const int flow_cap = k * k * l;
    const int layer_cap = k * k;
 
    auto a = [&](int i) { return 2 * l + 3 + 2 * i; };
    auto b = [&](int i) { return 2 * l + 4 + 2 * i; };
    auto t = [&](int i) { return 4 * l + 2 * k + 1 - i; };
 
    egraph<int>::value g(n, {}, true);
    g.edge_weighted();
 
    for (int i = 0; i + 1 < 2 * l - 1; ++i)
        g.add_edge(i, i + 1, flow_cap);
    for (int i = 0; i + 1 < 2 * l - 1; ++i)
        g.add_edge(t(i + 1), t(i), flow_cap);
 
    for (int i = 0; i < 2 * l - 1; i += 2) {
        g.add_edge(i, i % 4 == 0 ? p1 : p2, layer_cap);
        g.add_edge(i % 4 == 0 ? q1 : q2, t(i), layer_cap);
    }
 
    for (int i = 0; i < k; ++i) {
        g.add_edge(p1, a(i), flow_cap);
        g.add_edge(p2, b(i), flow_cap);
        g.add_edge(a(i), q2, flow_cap);
        g.add_edge(b(i), q1, flow_cap);
    }
 
    for (int i = 0; i < k; ++i)
        for (int j = 0; j < k; ++j)
            g.add_edge(a(i), b(j), 1);
 
    return g;
}
 
// Returns a mask of length 19938, with weights such that xor-ing with mt19937
// outputs yields 0.
// O(1).
template <typename T> std::vector<bool> mt19937_xor_hash() {
    static_assert(std::is_same_v<T, int> or std::is_same_v<T, long long>,
                  "hack: mt19937_xor_hash: T must be int or long long");
 
    constexpr std::size_t deg = 19937;
 
    std::bitset<deg + 1> a, b, c;
    b[deg] = c[deg] = 1;
    std::size_t l = 0, shift = 1;
    std::mt19937 rng32;
    std::mt19937_64 rng64;
    for (std::size_t n = 0; n < deg * 2; ++n) {
        a >>= 1;
        if constexpr (std::is_same_v<T, int>)
            a[deg] = rng32() & 1;
        else
            a[deg] = rng64() & 1;
 
        if ((c & a).count() % 2 == 0) {
            ++shift;
            continue;
        }
 
        std::bitset<deg + 1> oc = c;
        c ^= (b >> shift);
        if (2 * l <= n) {
            l = n + 1 - l;
            b = oc;
            shift = 1;
        } else {
            ++shift;
        }
    }
 
    std::vector<bool> mask(deg + 1);
    for (std::size_t i = 0; i <= deg; ++i)
        mask[i] = c[i];
    return mask;
}
 
// Convex polygon that breaks naive rotating calipers for maximum vertex
// distance (advances j while dist(i, next(j)) > dist(i, j) instead of using
// ccw).
// O(1).
inline std::vector<geometry::point<double>>
naive_rotating_calipers_max_dist_bug() {
    return {
        {-0.9846, -1.53251}, {0.49946, 1.19525},  {0.79916, 0.98291},
        {4.02136, -1.57843}, {3.92734, -2.37856}, {3.88558, -2.37188},
    };
}
 
namespace detail {
 
// Builds a hack block of order k (length fib(2k+1)).
// O(fib(2k+1)).
inline std::vector<int> segment_tree_beats_worst_case_block(int k) {
    tgen_ensure(k >= 1,
                "hack: segment_tree_beats_worst_case: k must be at least 1");
 
    std::vector<int> a(k + 1), b(k + 1);
    std::vector<std::vector<int>> vf(k + 1), vg(k + 1);
 
    a[1] = b[1] = 1;
    vf[1] = {1};
    vg[1] = {1, 0};
 
    for (int i = 2; i <= k; ++i) {
        b[i] = b[i - 1] + a[i - 1];
        a[i] = b[i] + a[i - 1];
        for (int x : vf[i - 1])
            vf[i].push_back(x + a[i] + b[i]);
        vf[i].push_back(a[i]);
        for (int x : vg[i - 1])
            vf[i].push_back(x + a[i]);
        vg[i] = vf[i];
        vg[i].push_back(0);
        for (int x : vg[i - 1])
            vg[i].push_back(x);
    }
 
    vf[k].push_back(0);
    return vf[k];
}
 
// Appends one update round for the tiled array (offset (round * an) mod L).
// O(fib(2k+1)).
inline void
segment_tree_beats_append_round(std::vector<std::vector<int>> &updates,
                                int block_len, int an, int bn, int n,
                                int round) {
    const int off = (round * an) % block_len;
    const int add_off = (off + block_len - bn) % block_len;
    for (int k = 0; k < block_len; ++k) {
        const int s = k * block_len * block_len;
        const int sub_end = off + an;
        if (sub_end <= block_len)
            updates.push_back({1, s + off, s + sub_end, bn});
        else {
            updates.push_back({1, s + off, s + block_len, bn});
            updates.push_back({1, s, s + (sub_end - block_len), bn});
        }
        const int add_end = add_off + bn;
        if (add_end <= block_len)
            updates.push_back({0, s + add_off, s + add_end, an});
        else {
            updates.push_back({0, s + add_off, s + block_len, an});
            updates.push_back({0, s, s + (add_end - block_len), an});
        }
    }
    updates.push_back({2, 0, n, an});
    for (int k = 0; k < block_len; ++k) {
        const int s = k * block_len * block_len;
        updates.push_back({3, s + (off + an - 1) % block_len, 0});
    }
}
 
} // namespace detail
 
// Array and updates for worst case of segment tree beats.
// O(fib(2k+1)^3 + q).
inline std::pair<std::vector<int>, std::vector<std::vector<int>>>
segment_tree_beats_worst_case(int k, int q) {
    tgen_ensure(k >= 1,
                "hack: segment_tree_beats_worst_case: k must be at least 1");
    tgen_ensure(k <= 7, "hack: segment_tree_beats_worst_case: k too large");
    tgen_ensure(q > 0,
                "hack: segment_tree_beats_worst_case: q must be positive");
 
    const auto &fib = math::fibonacci();
    const int block_len = fib[k * 2 + 1];
    const int an = fib[k * 2];
    const int bn = fib[k * 2 - 1];
 
    const int len = block_len;
    const int total = len * len * len;
 
    std::vector<int> block = detail::segment_tree_beats_worst_case_block(k);
    std::vector<int> arr(total, 0);
    for (int x = 0; x < block_len; ++x) {
        const int s = x * len * len;
        for (int i = 0; i < block_len; ++i)
            arr[s + i] = block[i];
    }
 
    std::vector<std::vector<int>> updates;
    updates.reserve(q);
    const int n = total;
    for (int round = 0; updates.size() < static_cast<std::size_t>(q); ++round) {
        detail::segment_tree_beats_append_round(updates, block_len, an, bn, n,
                                                round);
        if (updates.size() > static_cast<std::size_t>(q))
            updates.resize(q);
    }
    return {arr, updates};
}
 
} // namespace hack
 
/*********************
 *                   *
 *   MISCELLANEOUS   *
 *                   *
 *********************/
 
namespace misc {
 
// Generates a uniformly random balanced parentheses sequence with k '(' and k
// ')'. Valid means that for no prefix there are more ')' than '('.
// O(size).
inline std::string gen_parenthesis(int size) {
    tgen_ensure(size > 0 and size % 2 == 0,
                "misc: parenthesis: size must be a positive even number");
 
    int k = size / 2;
    std::string s;
    int open = 0, close = 0;
 
    for (int i = 0; i < size; ++i) {
        if (open == k) {
            s += ')';
            ++close;
            continue;
        }
        if (open == close) {
            s += '(';
            ++open;
            continue;
        }
 
        long long a = k - open, b = k - close, h = open - close;
 
        // Probability of placing '(':
        // P('(') = (k - open) * (h + 2) / ((k - open + k - close) * (h + 1))
        // Derived from ballot numbers ratio.
        long long num = a * (h + 2);
        long long den = (a + b) * (h + 1);
 
        if (next<long long>(1, den) <= num) {
            s += '(';
            ++open;
        } else {
            s += ')';
            ++close;
        }
    }
 
    return s;
}
 
} // namespace misc
 
} // namespace tgen