//! Bounded deterministic fuzzing helpers for the vendored Reed-Solomon implementation.

#[cfg(target_arch = "aarch64")]
use super::engine::Neon;
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
use super::engine::{Avx2, Avx512};
use super::{
    Decoder, Error, Plan,
    engine::{
        CANTOR_BASIS, DefaultEngine, Engine, GF_MODULUS, GF_ORDER, GF_POLYNOMIAL, GfElement, Naive,
        SHARD_CHUNK_BYTES, Scalar, ShardsRefMut,
    },
    rate::{DefaultRate, HighRate, LowRate, Rate, RateDecoder, RateEncoder},
};
use arbitrary::Unstructured;
use fixedbitset::FixedBitSet;

/// Shard lengths in bytes for rate cases: one 2-byte element, one 64-byte chunk, and one element
/// either side of one and two chunks.
const SHARD_SIZES: [usize; 6] = [2, 62, 64, 66, 126, 130];

/// Instantiate each supported engine, including implementations below the default CPU priority.
macro_rules! each_engine {
    ($runner:ident ( $($arg:expr),* )) => {{
        $runner::<Scalar>($($arg,)* Scalar::new);
        $runner::<DefaultEngine>($($arg,)* DefaultEngine::new);
        #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
        {
            if std::arch::is_x86_feature_detected!("avx512f")
                && std::arch::is_x86_feature_detected!("gfni")
            {
                $runner::<Avx512>($($arg,)* Avx512::new);
            }
            if std::arch::is_x86_feature_detected!("avx2") {
                $runner::<Avx2>($($arg,)* Avx2::new);
            }
        }
        #[cfg(target_arch = "aarch64")]
        if std::arch::is_aarch64_feature_detected!("neon") {
            $runner::<Neon>($($arg,)* Neon::new);
        }
    }};
}

/// Bounded comparisons of engines and engine-independent kernels against references.
#[derive(Clone, Copy, Debug, arbitrary::Arbitrary)]
pub enum EnginePlan {
    /// Compare multiplication, including empty and multi-chunk operands.
    Mul,
    /// Compare FFT and IFFT with offsets, truncation, and sentinel shards.
    Transform,
    /// Compare short Walsh locator convolution with the full-field evaluator.
    Locator,
    /// Compare the fused formal derivative with the unfused pass schedule.
    Derivative,
}

impl EnginePlan {
    /// Run the selected check.
    ///
    /// `Mul` and `Transform` compare every engine supported by this host. `Locator` and
    /// `Derivative` check engine-independent kernels once.
    pub fn run(self, u: &mut Unstructured<'_>) -> arbitrary::Result<()> {
        match self {
            Self::Mul => fuzz_mul(u),
            Self::Transform => fuzz_transform(u),
            Self::Locator => fuzz_locator(u),
            Self::Derivative => fuzz_derivative(u),
        }
    }
}

/// Checks multiplication on fuzzed chunks with `compare_mul`.
///
/// Draws `log_m` from logarithm edge cases, including `GF_MODULUS`, or at random. Draws 0, 1,
/// 2, 3, or 65 chunks, or 4 to 16.
fn fuzz_mul(u: &mut Unstructured<'_>) -> arbitrary::Result<()> {
    let log_m = match u.int_in_range(0..=7)? {
        0 => 0,
        1 => 1,
        2 => 2,
        3 => GF_MODULUS / 2,
        4 => GF_MODULUS / 2 + 1,
        5 => GF_MODULUS - 1,
        6 => GF_MODULUS,
        _ => u.arbitrary()?,
    };
    let mul_chunks = match u.int_in_range(0..=5)? {
        0 => 0,
        1 => 1,
        2 => 2,
        3 => 3,
        4 => 65,
        _ => u.int_in_range(4..=16)?,
    };
    let mut mul_input = vec![[0; SHARD_CHUNK_BYTES]; mul_chunks];
    let mut input = Input::new(u.bytes(u.len())?);
    fill_chunks(&mut mul_input, &mut input);
    compare_mul(&mul_input, log_m);
    Ok(())
}

/// Checks FFT and IFFT on a fuzzed block with `compare_transform`.
///
/// The block of `size = 2^k` shards, `k` in `0..=6`, starts at shard `pos` and is followed by up
/// to three suffix shards. `skew_delta` is an edge value or random in `0..=GF_ORDER - size`.
fn fuzz_transform(u: &mut Unstructured<'_>) -> arbitrary::Result<()> {
    let size = 1usize << u.int_in_range(0..=6)?;
    let pos = u.int_in_range(0..=5)?;
    let suffix = u.int_in_range(0..=3)?;
    let truncated_size = u.int_in_range(0..=size)?;
    let max_skew = GF_ORDER - size;
    let skew_delta = match u.int_in_range(0..=5)? {
        0 => 0,
        1 => 1.min(max_skew),
        2 => 7.min(max_skew),
        3 => max_skew,
        4 => 17.min(max_skew),
        _ => usize::from(u.arbitrary::<u16>()?) % (max_skew + 1),
    };
    let shard_chunks = match u.int_in_range(0..=5)? {
        0 => 0,
        1 => 1,
        2 => 2,
        3 => 17,
        4 => 65,
        _ => u.int_in_range(3..=8)?,
    };
    let shard_count = pos + size + suffix;
    let mut transform_input = vec![[0; SHARD_CHUNK_BYTES]; shard_count * shard_chunks];
    let mut input = Input::new(u.bytes(u.len())?);
    fill_chunks(&mut transform_input, &mut input);
    compare_transform(
        &transform_input,
        shard_count,
        shard_chunks,
        pos,
        size,
        truncated_size,
        skew_delta,
    );
    Ok(())
}

/// Checks the short Walsh locator on a fuzzed power-of-two domain `n <= 2^15` and truncation
/// `end` in `1..=n`.
fn fuzz_locator(u: &mut Unstructured<'_>) -> arbitrary::Result<()> {
    let n = 1usize << u.int_in_range(0..=15)?;
    let end = u.int_in_range(1..=n)?;
    compare_short_locator(n, end, u.arbitrary()?);
    Ok(())
}

/// Checks that `eval_poly_short` on `n` entries matches `Naive::eval_poly` modulo `GF_MODULUS`
/// on the first `n` outputs and never writes entry `n`.
///
/// Entries `0..end` hold values derived from `seed`, and the rest are zero. `n` must be a power
/// of two below `GF_ORDER`, and `end <= n`.
fn compare_short_locator(n: usize, end: usize, seed: u16) {
    // The convolution is linear modulo GF_MODULUS, so inputs need not be 0/1 flags. Mix zero,
    // GF_MODULUS (the other encoding of zero), and seeded residues.
    let mut short = [0; GF_ORDER];
    for (i, value) in short[..end].iter_mut().enumerate() {
        *value = match (i ^ usize::from(seed)) % 5 {
            0 => 0,
            1 => GF_MODULUS,
            _ => (i as u16).wrapping_mul(257) ^ seed,
        };
    }
    let mut full = short;
    Naive::eval_poly(&mut full, end);

    // The sentinel detects a write past the domain.
    short[n] = 0xa5a5;
    super::engine::utils::eval_poly_short(&mut short[..n], end);
    for i in 0..n {
        assert_eq!(
            short[i] % GF_MODULUS,
            full[i] % GF_MODULUS,
            "n={n} end={end} seed={seed} i={i}"
        );
    }
    assert_eq!(short[n], 0xa5a5, "n={n} end={end} wrote past the domain");
}

/// Shard counts for derivative cases. They bracket the four- and sixteen-shard leaf blocks and
/// each pass width.
const DERIVATIVE_COUNTS: [usize; 21] = [
    0, 1, 2, 3, 4, 5, 7, 8, 9, 15, 16, 17, 31, 32, 33, 63, 64, 65, 127, 128, 129,
];
/// Chunks per shard for derivative cases. They bracket the AVX-512 leaf's 8-chunk (512-byte)
/// minimum and its 64-chunk (4 KiB) exclusion.
const DERIVATIVE_CHUNKS: [usize; 10] = [0, 1, 3, 7, 8, 9, 17, 63, 64, 65];

/// Checks `formal_derivative` with `compare_derivative` on a fuzzed shard count and chunks per
/// shard.
///
/// The count comes from `DERIVATIVE_COUNTS` or from `0..=129`. The chunks per shard come from
/// `DERIVATIVE_CHUNKS`.
fn fuzz_derivative(u: &mut Unstructured<'_>) -> arbitrary::Result<()> {
    let count = if u.arbitrary::<bool>()? {
        DERIVATIVE_COUNTS[u.int_in_range(0..=20)?]
    } else {
        u.int_in_range(0..=129)?
    };
    let chunks = DERIVATIVE_CHUNKS[u.int_in_range(0..=9)?];
    let mut input = vec![[0; SHARD_CHUNK_BYTES]; count * chunks];
    fill_chunks(&mut input, &mut Input::new(u.bytes(u.len())?));
    compare_derivative(&input, count, chunks);
    Ok(())
}

/// Checks `formal_derivative` on `count` shards of `chunks` chunks each against the reference
/// pass schedule.
///
/// `input` holds the shards back to back, so its length must be `count * chunks`.
fn compare_derivative(input: &[[u8; SHARD_CHUNK_BYTES]], count: usize, chunks: usize) {
    // The reference runs the unclamped pass schedule over a zero-padded power-of-two block.
    // Padding shards only absorb higher padding shards and stay zero, so the first `count`
    // shards match the schedule with passes clamped at `count`.
    let padded_count = count.max(1).next_power_of_two();
    let mut expected = vec![[0; SHARD_CHUNK_BYTES]; padded_count * chunks];
    expected[..input.len()].copy_from_slice(input);
    let mut actual = input.to_vec();

    let mut reference = ShardsRefMut::new(padded_count, chunks, &mut expected);
    for i in 1..padded_count {
        let width = 1 << i.trailing_zeros();
        super::engine::utils::xor_within(&mut reference, i - width, i, width);
    }
    super::engine::utils::formal_derivative(&mut ShardsRefMut::new(count, chunks, &mut actual));
    assert_eq!(
        actual,
        expected[..input.len()],
        "count={count}, chunks={chunks}"
    );
}

/// Encoding rate exercised by the shared checks.
#[derive(Clone, Copy, Debug, arbitrary::Arbitrary)]
pub enum RateKind {
    /// More original shards than recovery shards.
    High,
    /// More recovery shards than original shards.
    Low,
    /// Switch between high and low rates when resetting.
    Default,
}

/// Bounded rate comparisons and backend-independent encoding/decoding contracts.
#[derive(Clone, Copy, Debug, arbitrary::Arbitrary)]
pub enum RatePlan {
    /// Compare encoding with the naive engine and decode its recovery shards.
    MatchesNaive(RateKind),
    /// Check roundtrips, automatic reuse, and reset with the same backend.
    Contract(RateKind),
    /// Check the public recovery decoder's reuse and missing recovery shards.
    Recovery,
    /// Compare prepared recovery against ordinary decoding across sizes, reuse, and mismatches.
    PreparedRecovery,
    /// Compare the rate-specific shortened locator against full-field arithmetic.
    Locator,
}

impl RatePlan {
    /// Run rate checks on every supported engine, exercise the public recovery decoder, or
    /// compare a rate's shortened erasure locator with the naive evaluator.
    pub fn run(self, u: &mut Unstructured<'_>) -> arbitrary::Result<()> {
        if matches!(self, Self::Locator) {
            return fuzz_rate_locator(u);
        }
        let small = u.int_in_range(2..=4)?;
        let large = u.int_in_range(9..=12)?;
        let shard_bytes_a = SHARD_SIZES[u.int_in_range(0..=5)?];
        let shard_bytes_b = SHARD_SIZES[u.int_in_range(0..=5)?];

        let (counts_a, counts_b) = match self {
            Self::MatchesNaive(RateKind::High) | Self::Contract(RateKind::High) => {
                ((large, small), (large - 1, small + 1))
            }
            Self::MatchesNaive(RateKind::Low) | Self::Contract(RateKind::Low) => {
                ((small, large), (small + 1, large - 1))
            }
            _ => ((large, small), (small, large)),
        };
        let mut input = Input::new(u.bytes(u.len())?);
        let case_a = RateCase::new(counts_a.0, counts_a.1, shard_bytes_a, &mut input);
        let case_b = RateCase::new(counts_b.0, counts_b.1, shard_bytes_b, &mut input);

        match self {
            Self::MatchesNaive(kind) => {
                each_engine!(rate_matches_naive(kind, &case_a, &case_b))
            }
            Self::Contract(kind) => {
                rate_contract::<Naive>(kind, &case_a, &case_b, Naive::new);
                each_engine!(rate_contract(kind, &case_a, &case_b));
            }
            Self::Recovery => exercise_recovery_reuse(&case_a, &case_b),
            Self::PreparedRecovery => exercise_prepared_reuse(&case_a, &case_b),
            Self::Locator => unreachable!("returned before case generation"),
        }
        Ok(())
    }
}

/// Checks a rate's shortened erasure locator with `compare_rate_locator` on fuzzed counts in
/// `1..=64`, a fuzzed rate, and a fuzzed received mask over the decoding domain.
fn fuzz_rate_locator(u: &mut Unstructured<'_>) -> arbitrary::Result<()> {
    let high = u.arbitrary::<bool>()?;
    let original_count = u.int_in_range(1usize..=64)?;
    let recovery_count = u.int_in_range(1usize..=64)?;
    let chunk = if high { recovery_count } else { original_count }.next_power_of_two();
    let end = chunk + if high { original_count } else { recovery_count };
    let mut received = FixedBitSet::with_capacity(end);
    for i in 0..end {
        if u.arbitrary::<bool>()? {
            received.insert(i);
        }
    }
    compare_rate_locator(high, original_count, recovery_count, &received);
    Ok(())
}

/// Checks a rate's shortened erasure locator against `Naive::eval_poly` over the full field.
///
/// `high` selects the high-rate layout (`rate_high::eval_erasures`) or the low-rate layout
/// (`rate_low::eval_walsh`). The decoding domain `0..end` places recovery shards (high rate) or
/// originals (low rate) from 0 and the other kind from `chunk`, the next power of two of the
/// first kind's count. `received` marks received positions in `0..end`. Only outputs below `end`
/// are compared, modulo `GF_MODULUS`.
fn compare_rate_locator(
    high: bool,
    original_count: usize,
    recovery_count: usize,
    received: &FixedBitSet,
) {
    // The reference marks erased positions over the full field. High rate places recovery
    // shards from 0 and originals from `chunk`, erases the padding between them, and truncates
    // at `end`. Low rate places originals from 0 and recovery shards from `chunk`, keeps the
    // padding known, and erases every position from `end`.
    let chunk = if high { recovery_count } else { original_count }.next_power_of_two();
    let end = chunk + if high { original_count } else { recovery_count };
    let mut expected = [0; GF_ORDER];
    if high {
        for i in 0..recovery_count {
            expected[i] = u16::from(!received[i]);
        }
        expected[recovery_count..chunk].fill(1);
        for i in chunk..end {
            expected[i] = u16::from(!received[i]);
        }
    } else {
        for i in 0..original_count {
            expected[i] = u16::from(!received[i]);
        }
        for i in chunk..end {
            expected[i] = u16::from(!received[i]);
        }
        expected[end..].fill(1);
    }
    Naive::eval_poly(&mut expected, if high { end } else { GF_ORDER });
    let mut actual = [0; GF_ORDER];
    let domain = &mut actual[..end.next_power_of_two()];
    if high {
        super::rate::rate_high::eval_erasures::<Scalar>(
            domain,
            original_count,
            recovery_count,
            received,
        );
    } else {
        super::rate::rate_low::eval_walsh::<Scalar>(
            domain,
            original_count,
            recovery_count,
            received,
        );
    }
    for i in 0..end {
        assert_eq!(
            actual[i] % GF_MODULUS,
            expected[i] % GF_MODULUS,
            "high={high} original={original_count} recovery={recovery_count} i={i}"
        );
    }
}

/// Cyclic byte source over fuzz input. Reads zeros when the input is empty.
struct Input<'a> {
    /// Fuzz bytes to cycle through.
    bytes: &'a [u8],
    /// Number of bytes read. The next read takes index `offset % bytes.len()`.
    offset: usize,
}

impl<'a> Input<'a> {
    /// Creates a source that starts at the first byte of `bytes`.
    const fn new(bytes: &'a [u8]) -> Self {
        Self { bytes, offset: 0 }
    }

    /// Returns the next byte, wrapping to the start after the last.
    ///
    /// Returns 0 if the input is empty.
    const fn byte(&mut self) -> u8 {
        if self.bytes.is_empty() {
            return 0;
        }
        let byte = self.bytes[self.offset % self.bytes.len()];
        self.offset += 1;
        byte
    }
}

/// Fills every chunk from `input`, adding chunk and lane offsets so the bytes still vary when a
/// short input repeats.
fn fill_chunks(chunks: &mut [[u8; SHARD_CHUNK_BYTES]], input: &mut Input<'_>) {
    for (chunk_index, chunk) in chunks.iter_mut().enumerate() {
        let (low, high) = chunk.split_at_mut(SHARD_CHUNK_BYTES / 2);
        for lane in 0..SHARD_CHUNK_BYTES / 2 {
            low[lane] = input
                .byte()
                .wrapping_add((chunk_index as u8).wrapping_mul(29))
                .wrapping_add(lane as u8);
            high[lane] = input
                .byte()
                .wrapping_add((chunk_index as u8).wrapping_mul(53))
                .wrapping_sub(lane as u8);
        }
    }
}

/// Checks `Naive` multiplication by the element with logarithm `log_m` against
/// `independent_mul`, then checks every engine against `Naive`.
///
/// Each chunk holds 32 elements, with low bytes in its first half and high bytes in its second.
fn compare_mul(input: &[[u8; SHARD_CHUNK_BYTES]], log_m: GfElement) {
    let mut expected = input.to_vec();
    for chunk in &mut expected {
        let (low, high) = chunk.split_at_mut(SHARD_CHUNK_BYTES / 2);
        for lane in 0..SHARD_CHUNK_BYTES / 2 {
            let value = GfElement::from(low[lane]) | (GfElement::from(high[lane]) << 8);
            let product = independent_mul(value, log_m);
            low[lane] = product as u8;
            high[lane] = (product >> 8) as u8;
        }
    }

    let mut naive = input.to_vec();
    Naive::new().mul(&mut naive, log_m);
    assert_eq!(naive, expected, "Naive mul differs for log {log_m}");

    each_engine!(check_mul(input, log_m, &naive));
}

/// Checks that engine `E` multiplies `input` by the element with logarithm `log_m` to give
/// `expected`.
fn check_mul<E: Engine>(
    input: &[[u8; SHARD_CHUNK_BYTES]],
    log_m: GfElement,
    expected: &[[u8; SHARD_CHUNK_BYTES]],
    new_engine: fn() -> E,
) {
    let mut actual = input.to_vec();
    new_engine().mul(&mut actual, log_m);
    assert_eq!(
        actual,
        expected,
        "{} mul differs for log {log_m}",
        core::any::type_name::<E>()
    );
}

/// Multiplies `value` by the element with logarithm `log_m` without the engine tables.
///
/// Elements are Cantor-basis coordinates. The oracle converts `value` to the polynomial basis,
/// multiplies by x^log_m reduced modulo GF_POLYNOMIAL, and converts back. It builds x^log_m by
/// square-and-multiply, so a table bug cannot cancel against the same bug in the engines.
fn independent_mul(value: GfElement, log_m: GfElement) -> GfElement {
    let mut coefficient = 1;
    let mut alpha = 2;
    let mut exponent = usize::from(log_m) % usize::from(GF_MODULUS);
    while exponent > 0 {
        if exponent & 1 == 1 {
            coefficient = polynomial_mul(coefficient, alpha);
        }
        alpha = polynomial_mul(alpha, alpha);
        exponent >>= 1;
    }
    independent_mul_coefficient(value, coefficient)
}

/// Multiplies Cantor-basis `value` by polynomial-basis `coefficient`.
fn independent_mul_coefficient(value: GfElement, coefficient: GfElement) -> GfElement {
    polynomial_to_cantor(polynomial_mul(cantor_to_polynomial(value), coefficient))
}

/// Converts Cantor-basis coordinates `value` to the polynomial basis.
///
/// Bit `i` of `value` selects `CANTOR_BASIS[i]`, written in the polynomial basis.
fn cantor_to_polynomial(value: GfElement) -> GfElement {
    let mut polynomial = 0;
    for (bit, basis) in CANTOR_BASIS.iter().enumerate() {
        if value & (1 << bit) != 0 {
            polynomial ^= basis;
        }
    }
    polynomial
}

/// Converts polynomial-basis `value` to Cantor-basis coordinates, inverting
/// `cantor_to_polynomial`.
fn polynomial_to_cantor(value: GfElement) -> GfElement {
    // Invert the basis change by Gauss-Jordan elimination over GF(2). Row `r` keeps
    // `values[r]` equal to the XOR of `CANTOR_BASIS[j]` over the set bits `j` of
    // `coordinates[r]`. Once `values[r] == 1 << r`, `coordinates[r]` holds the Cantor
    // coordinates of x^r. A pivot always exists because the basis is linearly independent.
    static INVERSE: std::sync::OnceLock<[GfElement; 16]> = std::sync::OnceLock::new();
    let inverse = INVERSE.get_or_init(|| {
        let mut values = CANTOR_BASIS;
        let mut coordinates = core::array::from_fn(|index| 1 << index);
        for pivot in 0..16 {
            let row = (pivot..16)
                .find(|row| values[*row] & (1 << pivot) != 0)
                .unwrap();
            values.swap(pivot, row);
            coordinates.swap(pivot, row);
            for row in 0..16 {
                if row != pivot && values[row] & (1 << pivot) != 0 {
                    values[row] ^= values[pivot];
                    coordinates[row] ^= coordinates[pivot];
                }
            }
        }
        debug_assert_eq!(values, core::array::from_fn(|index| 1 << index));
        coordinates
    });

    let mut coordinates = 0;
    for (bit, coordinate) in inverse.iter().enumerate() {
        if value & (1 << bit) != 0 {
            coordinates ^= coordinate;
        }
    }
    coordinates
}

/// Carry-less multiply, then reduce the degree-30 product modulo GF_POLYNOMIAL from the top bit.
fn polynomial_mul(left: GfElement, right: GfElement) -> GfElement {
    let mut product = 0u32;
    for bit in 0..16 {
        if right & (1 << bit) != 0 {
            product ^= u32::from(left) << bit;
        }
    }
    for bit in (16..=30).rev() {
        if product & (1 << bit) != 0 {
            product ^= (GF_POLYNOMIAL as u32) << (bit - 16);
        }
    }
    product as GfElement
}

/// Checks every engine's FFT and IFFT against `Naive` on shards `pos..pos + size` of `input`.
///
/// `input` holds `shard_count` shards of `shard_chunks` chunks each. Shards outside the block
/// must stay unchanged. FFT compares the first `truncated_size` block outputs. IFFT first zeroes
/// the block's input shards from `pos + truncated_size`, then compares the whole block.
///
/// # Panics
///
/// Panics if `size` is not a power of two at most `GF_ORDER`, `truncated_size > size`, the block
/// extends past `shard_count`, or `size > 1` and `skew_delta > GF_ORDER - size`.
fn compare_transform(
    input: &[[u8; SHARD_CHUNK_BYTES]],
    shard_count: usize,
    shard_chunks: usize,
    pos: usize,
    size: usize,
    truncated_size: usize,
    skew_delta: usize,
) {
    assert!(size.is_power_of_two());
    assert!(size <= GF_ORDER);
    assert!(truncated_size <= size);
    assert!(pos.checked_add(size).is_some_and(|end| end <= shard_count));
    assert!(size == 1 || skew_delta <= GF_ORDER - size);

    for inverse in [false, true] {
        let mut operation_input = input.to_vec();
        if inverse {
            for shard in pos + truncated_size..pos + size {
                operation_input[shard * shard_chunks..(shard + 1) * shard_chunks]
                    .fill([0; SHARD_CHUNK_BYTES]);
            }
        }

        let mut expected = operation_input.clone();
        apply_transform(
            &Naive::new(),
            &mut expected,
            shard_count,
            shard_chunks,
            pos,
            size,
            truncated_size,
            skew_delta,
            inverse,
        );

        for shard in 0..shard_count {
            if shard < pos || shard >= pos + size {
                let chunks = shard * shard_chunks..(shard + 1) * shard_chunks;
                assert_eq!(
                    expected[chunks.clone()],
                    input[chunks],
                    "Naive changed sentinel shard {shard}"
                );
            }
        }
        each_engine!(check_transform(
            &operation_input,
            input,
            &expected,
            shard_count,
            shard_chunks,
            pos,
            size,
            truncated_size,
            skew_delta,
            inverse
        ));
    }
}

/// Checks one FFT or IFFT of `operation_input` on engine `E` against `expected`.
///
/// `operation_input` is `input`, with the block's input suffix zeroed for IFFT. Shards outside the
/// block must match `input`. Block shards are compared where the transform specifies them, which
/// is the first `truncated_size` for FFT and the whole block for IFFT.
#[allow(clippy::too_many_arguments)]
fn check_transform<E: Engine>(
    operation_input: &[[u8; SHARD_CHUNK_BYTES]],
    input: &[[u8; SHARD_CHUNK_BYTES]],
    expected: &[[u8; SHARD_CHUNK_BYTES]],
    shard_count: usize,
    shard_chunks: usize,
    pos: usize,
    size: usize,
    truncated_size: usize,
    skew_delta: usize,
    inverse: bool,
    new_engine: fn() -> E,
) {
    let name = core::any::type_name::<E>();
    let mut actual = operation_input.to_vec();
    apply_transform(
        &new_engine(),
        &mut actual,
        shard_count,
        shard_chunks,
        pos,
        size,
        truncated_size,
        skew_delta,
        inverse,
    );
    for shard in 0..shard_count {
        let chunks = shard * shard_chunks..(shard + 1) * shard_chunks;
        if shard < pos || shard >= pos + size {
            assert_eq!(
                actual[chunks.clone()],
                input[chunks],
                "{name} changed sentinel shard {shard}"
            );
        } else if inverse || shard < pos + truncated_size {
            // FFT specifies a prefix. IFFT specifies the entire block after zeroing its input
            // suffix.
            assert_eq!(
                actual[chunks.clone()],
                expected[chunks],
                "{name} {} differs at shard {shard} for pos={pos} size={size} truncated={truncated_size} skew={skew_delta}",
                if inverse { "ifft" } else { "fft" }
            );
        }
    }
}

/// Runs the FFT of `engine`, or the IFFT if `inverse` is set, on shards `pos..pos + size`.
///
/// `data` holds `shard_count` shards of `shard_chunks` chunks each.
#[allow(clippy::too_many_arguments)]
fn apply_transform<E: Engine>(
    engine: &E,
    data: &mut [[u8; SHARD_CHUNK_BYTES]],
    shard_count: usize,
    shard_chunks: usize,
    pos: usize,
    size: usize,
    truncated_size: usize,
    skew_delta: usize,
    inverse: bool,
) {
    let mut shards = ShardsRefMut::new(shard_count, shard_chunks, data);
    if inverse {
        engine.ifft(&mut shards, pos, size, truncated_size, skew_delta);
    } else {
        engine.fft(&mut shards, pos, size, truncated_size, skew_delta);
    }
}

/// Rate encoding and decoding case with fuzzed shard contents and erasure offsets.
#[derive(Clone)]
struct RateCase {
    original_count: usize,
    recovery_count: usize,
    shard_bytes: usize,
    /// `original_count` original shards of `shard_bytes` bytes each.
    originals: Vec<Vec<u8>>,
    /// First original index dropped in the round with `start_delta` 0.
    original_start: usize,
    /// First recovery index provided in the round with `start_delta` 0.
    recovery_start: usize,
}

impl RateCase {
    /// Creates a case, drawing the erasure offsets and then the shard bytes from `input`.
    ///
    /// Both counts must be nonzero.
    fn new(
        original_count: usize,
        recovery_count: usize,
        shard_bytes: usize,
        input: &mut Input<'_>,
    ) -> Self {
        let original_start = usize::from(input.byte()) % original_count;
        let recovery_start = usize::from(input.byte()) % recovery_count;
        let mut originals = vec![vec![0; shard_bytes]; original_count];
        for (shard_index, shard) in originals.iter_mut().enumerate() {
            for (byte_index, byte) in shard.iter_mut().enumerate() {
                *byte = input
                    .byte()
                    .wrapping_add((shard_index as u8).wrapping_mul(31))
                    .wrapping_add((byte_index as u8).wrapping_mul(17));
            }
        }
        Self {
            original_count,
            recovery_count,
            shard_bytes,
            originals,
            original_start,
            recovery_start,
        }
    }

    /// Number of originals dropped and recovery shards provided in each round, so every
    /// decode receives exactly `original_count` shards. With at least two recovery shards, the
    /// cap leaves at least one recovery shard unprovided.
    fn missing_count(&self) -> usize {
        self.original_count.min(self.recovery_count - 1).max(1)
    }

    /// Returns a mask over originals that marks `missing_count()` consecutive indices as missing,
    /// starting at `original_start + start_delta` and wrapping at `original_count`.
    fn missing_originals(&self, start_delta: usize) -> Vec<bool> {
        let mut missing = vec![false; self.original_count];
        for offset in 0..self.missing_count() {
            missing[(self.original_start + start_delta + offset) % self.original_count] = true;
        }
        missing
    }

    /// Returns a mask over recovery shards that marks `missing_count()` consecutive indices as
    /// provided, starting at `recovery_start + start_delta` and wrapping at `recovery_count`.
    fn provided_recoveries(&self, start_delta: usize) -> Vec<bool> {
        let mut provided = vec![false; self.recovery_count];
        for offset in 0..self.missing_count() {
            provided[(self.recovery_start + start_delta + offset) % self.recovery_count] = true;
        }
        provided
    }
}

/// Encodes `case` with rate `R` on `engine` and returns the recovery shards.
fn encode<R, E>(case: &RateCase, engine: E) -> Vec<Vec<u8>>
where
    R: Rate<E>,
    E: Engine,
{
    let mut encoder = R::encoder(
        case.original_count,
        case.recovery_count,
        case.shard_bytes,
        engine,
        None,
    )
    .unwrap();
    encode_with::<R, E>(&mut encoder, case)
}

/// Adds every original of `case` to `encoder`, encodes, and returns the recovery shards.
///
/// `encoder` must be configured for the counts and shard size of `case`.
fn encode_with<R, E>(encoder: &mut R::RateEncoder, case: &RateCase) -> Vec<Vec<u8>>
where
    R: Rate<E>,
    E: Engine,
{
    for original in &case.originals {
        encoder.add_original_shard(original).unwrap();
    }
    encoder
        .encode()
        .unwrap()
        .recovery_iter()
        .map(<[u8]>::to_vec)
        .collect()
}

/// Runs one decode round and checks that the result holds exactly the missing originals.
///
/// Provides the originals and recovery shards selected by `start_delta`. `decoder` must be
/// configured for `case`, and `recovery` holds the recovery shards encoded from it.
fn decode_with<R, E>(
    decoder: &mut R::RateDecoder,
    case: &RateCase,
    recovery: &[Vec<u8>],
    start_delta: usize,
) where
    R: Rate<E>,
    E: Engine,
{
    let missing = case.missing_originals(start_delta);
    let provided_recovery = case.provided_recoveries(start_delta);
    for (index, original) in case.originals.iter().enumerate() {
        if !missing[index] {
            decoder.add_original_shard(index, original).unwrap();
        }
    }
    for (index, shard) in recovery.iter().enumerate() {
        if provided_recovery[index] {
            decoder.add_recovery_shard(index, shard).unwrap();
        }
    }

    let result = decoder.decode(false).unwrap().unwrap();
    for (index, original) in case.originals.iter().enumerate() {
        if missing[index] {
            assert_eq!(result.original(index), Some(original.as_slice()));
        } else {
            assert_eq!(result.original(index), None);
        }
    }
}

/// Checks encode and decode roundtrips with rate `R` on one engine, including reuse and reset.
///
/// The encoder encodes `case_a`, resets, and encodes `case_b`. The decoder decodes two rounds of
/// `case_a` without a reset, then resets and decodes two rounds of `case_b`.
fn check_rate_contract<R, E>(case_a: &RateCase, case_b: &RateCase, new_engine: fn() -> E)
where
    R: Rate<E>,
    E: Engine,
{
    let mut encoder = R::encoder(
        case_a.original_count,
        case_a.recovery_count,
        case_a.shard_bytes,
        new_engine(),
        None,
    )
    .unwrap();
    let recovery_a = encode_with::<R, E>(&mut encoder, case_a);
    encoder
        .reset(
            case_b.original_count,
            case_b.recovery_count,
            case_b.shard_bytes,
        )
        .unwrap();
    let recovery_b = encode_with::<R, E>(&mut encoder, case_b);

    let mut decoder = R::decoder(
        case_a.original_count,
        case_a.recovery_count,
        case_a.shard_bytes,
        new_engine(),
        None,
    )
    .unwrap();
    decode_with::<R, E>(&mut decoder, case_a, &recovery_a, 0);
    decode_with::<R, E>(&mut decoder, case_a, &recovery_a, 1);
    decoder
        .reset(
            case_b.original_count,
            case_b.recovery_count,
            case_b.shard_bytes,
        )
        .unwrap();
    decode_with::<R, E>(&mut decoder, case_b, &recovery_b, 1);
    decode_with::<R, E>(&mut decoder, case_b, &recovery_b, 0);
}

/// Checks that rate `R` decodes `case_a` from the reference recovery shards `expected_a`, then
/// resets and decodes `case_b` from `expected_b`.
fn decode_reference<R, E>(
    case_a: &RateCase,
    case_b: &RateCase,
    expected_a: &[Vec<u8>],
    expected_b: &[Vec<u8>],
    new_engine: fn() -> E,
) where
    R: Rate<E>,
    E: Engine,
{
    let mut decoder = R::decoder(
        case_a.original_count,
        case_a.recovery_count,
        case_a.shard_bytes,
        new_engine(),
        None,
    )
    .unwrap();
    decode_with::<R, E>(&mut decoder, case_a, expected_a, 0);
    decoder
        .reset(
            case_b.original_count,
            case_b.recovery_count,
            case_b.shard_bytes,
        )
        .unwrap();
    decode_with::<R, E>(&mut decoder, case_b, expected_b, 1);
}

/// Runs `check_rate_contract` on engine `E` with the rate selected by `kind`.
fn rate_contract<E: Engine>(
    kind: RateKind,
    case_a: &RateCase,
    case_b: &RateCase,
    new_engine: fn() -> E,
) {
    match kind {
        RateKind::High => check_rate_contract::<HighRate<E>, E>(case_a, case_b, new_engine),
        RateKind::Low => check_rate_contract::<LowRate<E>, E>(case_a, case_b, new_engine),
        RateKind::Default => check_rate_contract::<DefaultRate<E>, E>(case_a, case_b, new_engine),
    }
}

/// Runs `compare_rate` on engine `E` with the rate selected by `kind`.
fn rate_matches_naive<E: Engine>(
    kind: RateKind,
    case_a: &RateCase,
    case_b: &RateCase,
    new_engine: fn() -> E,
) {
    match kind {
        RateKind::High => {
            compare_rate::<HighRate<E>, HighRate<Naive>, E>(case_a, case_b, new_engine)
        }
        RateKind::Low => compare_rate::<LowRate<E>, LowRate<Naive>, E>(case_a, case_b, new_engine),
        RateKind::Default => {
            compare_rate::<DefaultRate<E>, DefaultRate<Naive>, E>(case_a, case_b, new_engine)
        }
    }
}

/// Checks that rate `R` on engine `E` encodes `case_a` and, after a reset, `case_b` to the same
/// recovery shards as `Reference` on `Naive`, then decodes both with `decode_reference`.
fn compare_rate<R, Reference, E>(case_a: &RateCase, case_b: &RateCase, new_engine: fn() -> E)
where
    R: Rate<E>,
    Reference: Rate<Naive>,
    E: Engine,
{
    let expected_a = encode::<Reference, _>(case_a, Naive::new());
    let expected_b = encode::<Reference, _>(case_b, Naive::new());
    let mut encoder = R::encoder(
        case_a.original_count,
        case_a.recovery_count,
        case_a.shard_bytes,
        new_engine(),
        None,
    )
    .unwrap();
    assert_eq!(encode_with::<R, E>(&mut encoder, case_a), expected_a);
    encoder
        .reset(
            case_b.original_count,
            case_b.recovery_count,
            case_b.shard_bytes,
        )
        .unwrap();
    assert_eq!(encode_with::<R, E>(&mut encoder, case_b), expected_b);
    decode_reference::<R, E>(case_a, case_b, &expected_a, &expected_b, new_engine);
}

/// Runs one `decode_with_recovery` round on the public `Decoder` and checks that it restores
/// exactly the missing originals and the unprovided recovery shards.
///
/// `start_delta` selects the round's erasures. `recovery` holds the recovery shards of `case`.
fn recovery_round(
    decoder: &mut Decoder,
    case: &RateCase,
    recovery: &[Vec<u8>],
    start_delta: usize,
) {
    let missing = case.missing_originals(start_delta);
    let provided_recovery = case.provided_recoveries(start_delta);
    for (index, original) in case.originals.iter().enumerate() {
        if !missing[index] {
            decoder.add_original_shard(index, original).unwrap();
        }
    }
    for (index, shard) in recovery.iter().enumerate() {
        if provided_recovery[index] {
            decoder.add_recovery_shard(index, shard).unwrap();
        }
    }

    let result = decoder.decode_with_recovery().unwrap().unwrap();
    for (index, original) in case.originals.iter().enumerate() {
        let expected = missing[index].then_some(original.as_slice());
        assert_eq!(result.original(index), expected);
    }
    for (index, shard) in recovery.iter().enumerate() {
        let expected = (!provided_recovery[index]).then_some(shard.as_slice());
        assert_eq!(result.recovery(index), expected);
    }
}

/// Checks the public `Decoder` over two rounds of `case_a` without a reset, then one round of
/// `case_b` after a reset.
fn exercise_recovery_reuse(case_a: &RateCase, case_b: &RateCase) {
    let expected_a = encode::<DefaultRate<Naive>, _>(case_a, Naive::new());
    let expected_b = encode::<DefaultRate<Naive>, _>(case_b, Naive::new());
    let mut decoder = Decoder::new(
        case_a.original_count,
        case_a.recovery_count,
        case_a.shard_bytes,
    )
    .unwrap();

    recovery_round(&mut decoder, case_a, &expected_a, 0);
    recovery_round(&mut decoder, case_a, &expected_a, 1);
    decoder
        .reset(
            case_b.original_count,
            case_b.recovery_count,
            case_b.shard_bytes,
        )
        .unwrap();
    recovery_round(&mut decoder, case_b, &expected_b, 0);
}

/// Builds the `Plan` for the erasures of round `start_delta` and runs
/// `prepared_round_with_plan` with it.
fn prepared_round(
    decoder: &mut Decoder,
    case: &RateCase,
    recovery: &[Vec<u8>],
    start_delta: usize,
) {
    let missing = case.missing_originals(start_delta);
    let provided = case.provided_recoveries(start_delta);
    let originals = (0..case.original_count).filter(|&i| !missing[i]);
    let recoveries = (0..case.recovery_count).filter(|&i| provided[i]);
    let plan = Plan::new(
        case.original_count,
        case.recovery_count,
        originals,
        recoveries,
    )
    .unwrap();
    prepared_round_with_plan(decoder, case, recovery, start_delta, &plan);
}

/// Runs one prepared decode round with `plan` and checks it against ordinary decoding of the
/// same shards.
///
/// First checks that a plan with a different original count, a plan with one more recovery
/// index, and a plan with one provided recovery index swapped for an unprovided one each return
/// `Error::PlanMismatch` and leave the submitted shards usable. `case` must have at least two
/// recovery shards, so one stays unprovided for the second and third plans.
fn prepared_round_with_plan(
    decoder: &mut Decoder,
    case: &RateCase,
    recovery: &[Vec<u8>],
    start_delta: usize,
    plan: &Plan,
) {
    let missing = case.missing_originals(start_delta);
    let provided = case.provided_recoveries(start_delta);
    let mut reference =
        Decoder::new(case.original_count, case.recovery_count, case.shard_bytes).unwrap();
    for (index, shard) in case.originals.iter().enumerate() {
        if !missing[index] {
            decoder.add_original_shard(index, shard).unwrap();
            reference.add_original_shard(index, shard).unwrap();
        }
    }
    for (index, shard) in recovery.iter().enumerate() {
        if provided[index] {
            decoder.add_recovery_shard(index, shard).unwrap();
            reference.add_recovery_shard(index, shard).unwrap();
        }
    }

    // A wrong count and a changed mask must leave the submitted shards usable for a retry.
    let other_counts = Plan::new(
        case.original_count + 1,
        case.recovery_count,
        0..case.original_count + 1,
        [],
    )
    .unwrap();
    assert!(matches!(
        decoder.decode_with_plan(&other_counts),
        Err(Error::PlanMismatch)
    ));
    let additional = provided.iter().position(|present| !present).unwrap();
    let changed = Plan::new(
        case.original_count,
        case.recovery_count,
        (0..case.original_count).filter(|&i| !missing[i]),
        (0..case.recovery_count).filter(|&i| provided[i] || i == additional),
    )
    .unwrap();
    assert!(matches!(
        decoder.decode_with_recovery_plan(&changed),
        Err(Error::PlanMismatch)
    ));

    // A plan with the same counts and one provided recovery index swapped for an unprovided one
    // must also mismatch.
    let swapped = provided.iter().position(|&present| present).unwrap();
    let exchanged = Plan::new(
        case.original_count,
        case.recovery_count,
        (0..case.original_count).filter(|&i| !missing[i]),
        (0..case.recovery_count).filter(|&i| (provided[i] && i != swapped) || i == additional),
    )
    .unwrap();
    assert!(matches!(
        decoder.decode_with_recovery_plan(&exchanged),
        Err(Error::PlanMismatch)
    ));
    let result = decoder.decode_with_recovery_plan(plan).unwrap().unwrap();
    let ordinary = reference.decode_with_recovery().unwrap().unwrap();
    for (index, shard) in case.originals.iter().enumerate() {
        assert_eq!(result.original(index), ordinary.original(index));
        assert_eq!(
            result.original(index),
            missing[index].then_some(shard.as_slice())
        );
    }
    for (index, shard) in recovery.iter().enumerate() {
        assert_eq!(result.recovery(index), ordinary.recovery(index));
        assert_eq!(
            result.recovery(index),
            (!provided[index]).then_some(shard.as_slice())
        );
    }
}

/// Checks prepared recovery across decoder reuse: two rounds of `case_a`, a plan reused for
/// 2-byte shards, a plan with every original present, and two rounds of `case_b` after a reset.
fn exercise_prepared_reuse(case_a: &RateCase, case_b: &RateCase) {
    let recovery_a = encode::<DefaultRate<Naive>, _>(case_a, Naive::new());
    let recovery_b = encode::<DefaultRate<Naive>, _>(case_b, Naive::new());
    let mut decoder = Decoder::new(
        case_a.original_count,
        case_a.recovery_count,
        case_a.shard_bytes,
    )
    .unwrap();

    // Decode two received masks in sequence without a reset.
    prepared_round(&mut decoder, case_a, &recovery_a, 0);
    prepared_round(&mut decoder, case_a, &recovery_a, 1);

    // A plan does not depend on shard bytes. Reuse the first mask's plan for 2-byte shards.
    let missing = case_a.missing_originals(0);
    let provided = case_a.provided_recoveries(0);
    let shared = Plan::new(
        case_a.original_count,
        case_a.recovery_count,
        (0..case_a.original_count).filter(|&i| !missing[i]),
        (0..case_a.recovery_count).filter(|&i| provided[i]),
    )
    .unwrap();
    let mut short = case_a.clone();
    short.shard_bytes = 2;
    for original in &mut short.originals {
        original.truncate(2);
    }
    let short_recovery = encode::<DefaultRate<Naive>, _>(&short, Naive::new());
    decoder
        .reset(short.original_count, short.recovery_count, 2)
        .unwrap();
    prepared_round_with_plan(&mut decoder, &short, &short_recovery, 0, &shared);

    // A matching plan with every original present returns None.
    decoder
        .reset(
            case_a.original_count,
            case_a.recovery_count,
            case_a.shard_bytes,
        )
        .unwrap();
    let all = Plan::new(
        case_a.original_count,
        case_a.recovery_count,
        0..case_a.original_count,
        [],
    )
    .unwrap();
    for (i, original) in case_a.originals.iter().enumerate() {
        decoder.add_original_shard(i, original).unwrap();
    }
    assert!(decoder.decode_with_recovery_plan(&all).unwrap().is_none());

    // Reset to the second shape and decode both of its masks.
    decoder
        .reset(
            case_b.original_count,
            case_b.recovery_count,
            case_b.shard_bytes,
        )
        .unwrap();
    prepared_round(&mut decoder, case_b, &recovery_b, 0);
    prepared_round(&mut decoder, case_b, &recovery_b, 1);
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::reed_solomon::{engine::GF_BITS, rate::rate_low::DIRECT_EVALUATION_LIMIT};
    use commonware_invariants::minifuzz;

    /// Builds a `RateCase` whose shard bytes and erasure offsets derive from `seed`.
    fn fixed_case(
        original_count: usize,
        recovery_count: usize,
        shard_bytes: usize,
        seed: u8,
    ) -> RateCase {
        RateCase::new(
            original_count,
            recovery_count,
            shard_bytes,
            &mut Input::new(&[seed, seed.wrapping_mul(17), seed.wrapping_add(91)]),
        )
    }

    /// Checks that `E::eval_poly` matches `Naive` on `input` truncated at `truncated_size`.
    ///
    /// `_new_engine` only fits the `each_engine!` calling convention.
    fn compare_eval_poly<E: Engine>(
        input: &[GfElement; GF_ORDER],
        truncated_size: usize,
        _new_engine: fn() -> E,
    ) {
        let mut expected: Box<[GfElement; GF_ORDER]> =
            input.to_vec().into_boxed_slice().try_into().unwrap();
        Naive::eval_poly(expected.as_mut(), truncated_size);
        let mut actual: Box<[GfElement; GF_ORDER]> =
            input.to_vec().into_boxed_slice().try_into().unwrap();
        E::eval_poly(actual.as_mut(), truncated_size);
        assert_eq!(actual, expected);
    }

    #[test]
    fn minifuzz_derivative() {
        minifuzz::Builder::default()
            .with_seed(0)
            .with_search_limit(64)
            .test(|u| EnginePlan::Derivative.run(u));
    }

    #[test]
    fn derivative_boundary_schedule() {
        // Chunk counts around 8 and 64 straddle the AVX-512 leaf guards, which Miri never
        // takes. Miri keeps every shard count with fewer chunk counts.
        let chunk_counts: &[usize] = if cfg!(miri) {
            &[0, 1, 8]
        } else {
            &[0, 1, 7, 8, 9, 63, 64, 65]
        };
        for count in [0, 1, 3, 4, 5, 15, 16, 17, 31, 32, 33, 64, 65] {
            for &chunks in chunk_counts {
                let mut input = vec![[0; SHARD_CHUNK_BYTES]; count * chunks];
                fill_chunks(&mut input, &mut Input::new(&[3, 0xff, 19, 0x80]));
                compare_derivative(&input, count, chunks);
            }
        }
    }

    #[test]
    fn minifuzz_mul() {
        minifuzz::Builder::default()
            .with_seed(0)
            .with_search_limit(64)
            .test(|u| EnginePlan::Mul.run(u));
    }

    #[test]
    fn minifuzz_transform() {
        minifuzz::Builder::default()
            .with_seed(0)
            .with_search_limit(64)
            .test(|u| EnginePlan::Transform.run(u));
    }

    #[test]
    fn minifuzz_engine_locator() {
        minifuzz::Builder::default()
            .with_seed(0)
            .with_search_limit(32)
            .test(|u| EnginePlan::Locator.run(u));
    }

    #[test]
    fn locator_every_short_level() {
        // Build each cached short Walsh kernel and check it at the domain edges.
        for bits in 0..GF_BITS {
            let n = 1usize << bits;
            for end in [1, n / 2 + 1, n] {
                compare_short_locator(n, end, 0x5a5a);
            }
        }
    }

    #[test]
    fn minifuzz_rate_locator() {
        minifuzz::Builder::default()
            .with_seed(0)
            .with_search_limit(32)
            .test(|u| RatePlan::Locator.run(u));
    }

    #[test]
    fn locator_domain_boundaries() {
        // Straddle several Walsh size changes and the switch to the full field above 32768. The
        // low-rate check calls `eval_walsh` directly, which ignores DIRECT_EVALUATION_LIMIT.
        for high in [false, true] {
            for end in [
                2, 3, 4, 7, 8, 15, 16, 127, 128, 129, 511, 512, 513, 32767, 32768, 32769, 65535,
                65536,
            ] {
                // Chunk size one places the other group immediately after it.
                let (original_count, recovery_count) =
                    if high { (end - 1, 1) } else { (1, end - 1) };
                for pattern in 0..3 {
                    let mut received = FixedBitSet::with_capacity(end);
                    for i in 0..end {
                        if pattern == 0 || (pattern == 1 && i % 3 != 0) {
                            received.insert(i);
                        }
                    }
                    compare_rate_locator(high, original_count, recovery_count, &received);
                }
            }
        }
    }

    #[test]
    fn minifuzz_rate_matches_naive() {
        for kind in [RateKind::High, RateKind::Low, RateKind::Default] {
            minifuzz::Builder::default()
                .with_seed(0)
                .with_search_limit(32)
                .test(|u| RatePlan::MatchesNaive(kind).run(u));
        }
    }

    #[test]
    fn minifuzz_rate_contract() {
        for kind in [RateKind::High, RateKind::Low, RateKind::Default] {
            minifuzz::Builder::default()
                .with_seed(0)
                .with_search_limit(32)
                .test(|u| RatePlan::Contract(kind).run(u));
        }
    }

    #[test]
    fn minifuzz_recovery() {
        minifuzz::Builder::default()
            .with_seed(0)
            .with_search_limit(32)
            .test(|u| RatePlan::Recovery.run(u));
    }

    #[test]
    fn minifuzz_prepared_recovery() {
        minifuzz::Builder::default()
            .with_seed(0)
            .with_search_limit(32)
            .test(|u| RatePlan::PreparedRecovery.run(u));
    }

    #[test]
    fn prepared_received_mask_reuse() {
        // Reuse larger masks for both partial and full final words in each rate.
        for (original_count, recovery_count) in [(3, 5), (5, 3), (16, 48), (48, 16)] {
            let large = fixed_case(original_count * 4, recovery_count * 4, 2, 21);
            let small = fixed_case(original_count, recovery_count, 2, 23);
            exercise_prepared_reuse(&large, &small);
        }
    }

    #[test]
    fn prepared_low_locator_boundaries() {
        let high = fixed_case(11, 3, 2, 23);

        // Three originals pad to a chunk of 4, so the decoding domain ends at 4 + recovery_count.
        // Straddle DIRECT_EVALUATION_LIMIT, then take the transform path with a larger domain.
        let limit = DIRECT_EVALUATION_LIMIT - 4;
        for recovery_count in [limit - 1, limit, limit + 1, 510] {
            let low = fixed_case(3, recovery_count, 66, 21);
            exercise_prepared_reuse(&low, &high);
        }
    }

    #[test]
    fn engines_match_naive_on_field_and_log_edges() {
        let mut chunks = vec![[0; SHARD_CHUNK_BYTES]; 3];
        for (chunk_index, chunk) in chunks.iter_mut().enumerate() {
            let (low, high) = chunk.split_at_mut(SHARD_CHUNK_BYTES / 2);
            for lane in 0..SHARD_CHUNK_BYTES / 2 {
                let value = match (chunk_index * 32 + lane) % 8 {
                    0 => 0,
                    1 => 1,
                    2 => 0x00ff,
                    3 => 0x0100,
                    4 => 0x7fff,
                    5 => 0x8000,
                    6 => 0xfffe,
                    _ => 0xffff,
                };
                low[lane] = value as u8;
                high[lane] = (value >> 8) as u8;
            }
        }
        for log_m in [0, 1, 2, 0x7fff, 0x8000, 0xfffe, 0xffff] {
            compare_mul(&[], log_m);
            compare_mul(&chunks, log_m);
        }
        let mut multi_vector = vec![[0; SHARD_CHUNK_BYTES]; 65];
        fill_chunks(&mut multi_vector, &mut Input::new(&[3, 0xff, 19, 0x80, 0]));
        compare_mul(&multi_vector, 12_345);
    }

    #[test]
    fn log_tables_match_independent_field_multiplication() {
        let tables = super::super::engine::tables::get_exp_log();
        let mut coefficient = 1;
        for log_m in 0..=GF_MODULUS {
            if log_m == GF_MODULUS {
                assert_eq!(coefficient, 1, "generator did not close at GF_MODULUS");
            } else if log_m > 0 {
                assert_ne!(coefficient, 1, "generator cycled early at log {log_m}");
            }
            for shift in [0, 4, 8, 12] {
                let bit_products: [GfElement; 4] = core::array::from_fn(|bit| {
                    independent_mul_coefficient(1 << (shift + bit), coefficient)
                });
                for nibble in 0..16 {
                    let value = nibble << shift;
                    let expected = bit_products
                        .iter()
                        .enumerate()
                        .filter(|(bit, _)| nibble & (1 << bit) != 0)
                        .fold(0, |product, (_, bit_product)| product ^ bit_product);
                    assert_eq!(
                        super::super::engine::tables::mul(value, log_m, &tables.exp, &tables.log,),
                        expected,
                        "table mul differs for value={value} log={log_m}",
                    );
                }
            }
            coefficient = polynomial_mul(coefficient, 2);
        }
    }

    #[test]
    fn engines_match_naive_on_offset_truncated_and_skewed_transforms() {
        // Size-4 cases pick butterfly branches through `skew_delta`. The transform reads
        // `log_m01`, `log_m02`, and `log_m23` from `skew[skew_delta..skew_delta + 3]`, and
        // `skew` holds `GF_MODULUS` (a zero coefficient) exactly at indices `2^k - 1`.
        // `skew_delta` 4 zeroes none and reaches the fused AVX-512 path. 0 zeroes `log_m01`
        // and `log_m02`, 1 zeroes `log_m01` and `log_m23`, 2 zeroes `log_m02`, and 3 and 7
        // zero `log_m01`.
        for (pos, size, truncated_size, skew_delta, shard_chunks) in [
            (0, 1, 0, GF_ORDER - 1, 1),
            (1, 2, 1, 0, 2),
            (2, 4, 1, 3, 2),
            (3, 4, 3, 7, 1),
            (2, 8, 5, 31, 3),
            (5, 16, 16, GF_ORDER - 16, 2),
            (1, 4, 4, 4, 0),
            (1, 4, 0, 4, 1),
            (1, 4, 4, 0, 1),
            (2, 4, 3, 1, 65),
            (3, 4, 2, 2, 1),
            (2, 4, 4, 4, 65),
            (1, 16, 13, 0, 1),
            (2, 64, 37, 17, 17),
        ] {
            let shard_count = pos + size + 2;
            let mut input = vec![[0; SHARD_CHUNK_BYTES]; shard_count * shard_chunks];
            fill_chunks(&mut input, &mut Input::new(&[3, 0xff, 19, 0x80, 0]));
            compare_transform(
                &input,
                shard_count,
                shard_chunks,
                pos,
                size,
                truncated_size,
                skew_delta,
            );
        }
    }

    #[test]
    fn maximum_domain_transform_regression() {
        let mut input = vec![[0; SHARD_CHUNK_BYTES]; GF_ORDER];
        fill_chunks(&mut input, &mut Input::new(&[3, 0xff, 19, 0x80, 0]));
        compare_transform(&input, GF_ORDER, 1, 0, GF_ORDER, 1, 0);
    }

    #[test]
    fn optimized_eval_poly_matches_naive() {
        for truncated_size in [0, 1, 257] {
            let mut input: Box<[GfElement; GF_ORDER]> =
                vec![0; GF_ORDER].into_boxed_slice().try_into().unwrap();
            for (index, value) in input.iter_mut().enumerate().take(truncated_size) {
                *value = [0, 1, 2, 0x7fff, 0x8000, 0xfffe, 0xffff][index % 7];
            }
            each_engine!(compare_eval_poly(&input, truncated_size));
        }
    }

    #[test]
    fn rate_engines_match_naive_with_partial_chunks_and_mixed_erasures() {
        let high_a = fixed_case(11, 3, 66, 3);
        let high_b = fixed_case(9, 4, 130, 5);
        each_engine!(rate_matches_naive(RateKind::High, &high_a, &high_b));

        let low_a = fixed_case(3, 11, 126, 7);
        let low_b = fixed_case(4, 9, 62, 11);
        each_engine!(rate_matches_naive(RateKind::Low, &low_a, &low_b));
    }

    #[test]
    fn default_rate_switches_and_recovery_decoder_reuses_work() {
        let high = fixed_case(11, 3, 66, 13);
        let low = fixed_case(3, 11, 130, 17);
        each_engine!(rate_matches_naive(RateKind::Default, &high, &low));
        exercise_recovery_reuse(&high, &low);
    }
}