rustfs_erasure_codec/core/

stream.rs

//! Streaming encode/verify/reconstruct for large data.
//!
//! Provides [`encode_stream`](crate::galois_8::ReedSolomon::encode_stream),
//! [`verify_stream`](crate::galois_8::ReedSolomon::verify_stream), and
//! [`reconstruct_stream`](crate::galois_8::ReedSolomon::reconstruct_stream) methods that
//! process data in configurable blocks, avoiding the need to load entire
//! files into memory.
//!
//! # Example
//!
//! ```no_run
//! use rustfs_erasure_codec::galois_8::ReedSolomon;
//! use rustfs_erasure_codec::stream::StreamOptions;
//! use std::fs::File;
//! use std::io::BufReader;
//!
//! let rs = ReedSolomon::new(10, 4).unwrap();
//!
//! let mut data_readers: Vec<BufReader<File>> = (0..10)
//!     .map(|i| BufReader::new(File::open(format!("data_{i}.bin")).unwrap()))
//!     .collect();
//! let mut parity_writers: Vec<File> = (0..4)
//!     .map(|i| File::create(format!("parity_{i}.bin")).unwrap())
//!     .collect();
//!
//! let opts = StreamOptions::new().with_block_size(4 * 1024 * 1024);
//! rs.encode_stream(&mut data_readers, &mut parity_writers, &opts).unwrap();
//! ```

use std::io::Error;
use std::mem;

/// Lower bound for the read/write block size (1 KiB).
const MIN_BLOCK_SIZE_BYTES: usize = 1024;
/// Upper bound for the read/write block size (16 MiB).
const MAX_BLOCK_SIZE_BYTES: usize = 16 * 1024 * 1024;

// ---------------------------------------------------------------------------
// StreamOptions
// ---------------------------------------------------------------------------

/// I/O scheduling mode for streaming operations.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum StreamIoMode {
    /// Choose serial or parallel I/O using conservative built-in thresholds.
    Auto,
    /// Read and write shard streams serially.
    Serial,
    /// Read and write shard streams with rayon parallel iterators.
    Parallel,
}

/// Configuration for streaming encode/verify/reconstruct operations.
#[derive(Debug, Clone)]
pub struct StreamOptions {
    /// Block size in bytes for each read/write cycle. Default: 4 MiB.
    ///
    /// Larger blocks reduce syscall overhead but increase memory usage.
    /// Recommended range: 256 KiB – 16 MiB.
    pub block_size: usize,
    /// I/O scheduling mode for stream reads and writes. Default: Auto.
    pub io_mode: StreamIoMode,
}

impl Default for StreamOptions {
    fn default() -> Self {
        Self {
            block_size: 4 * 1024 * 1024,
            io_mode: StreamIoMode::Auto,
        }
    }
}

impl StreamOptions {
    /// Create a new `StreamOptions` with default settings.
    pub fn new() -> Self {
        Self::default()
    }

    /// Set the block size (clamped to `[1 KiB, 16 MiB]`).
    pub fn with_block_size(mut self, size: usize) -> Self {
        self.block_size = size.clamp(MIN_BLOCK_SIZE_BYTES, MAX_BLOCK_SIZE_BYTES);
        self
    }

    /// Set the stream I/O scheduling mode.
    pub fn with_io_mode(mut self, mode: StreamIoMode) -> Self {
        self.io_mode = mode;
        self
    }
}

// ---------------------------------------------------------------------------
// StreamError
// ---------------------------------------------------------------------------

/// Error returned by streaming operations.
#[derive(Debug)]
pub struct StreamError {
    /// Index of the shard that caused the error.
    pub shard_index: usize,
    /// The kind of error.
    pub kind: StreamErrorKind,
}

/// Classification of [`StreamError`].
#[derive(Debug)]
pub enum StreamErrorKind {
    /// I/O error while reading a shard.
    Read(std::io::Error),
    /// I/O error while writing a shard.
    Write(std::io::Error),
    /// Error from the underlying codec (encode/verify/reconstruct).
    Codec(crate::Error),
}

impl core::fmt::Display for StreamError {
    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
        match &self.kind {
            StreamErrorKind::Read(e) => {
                write!(f, "read error on shard {}: {}", self.shard_index, e)
            }
            StreamErrorKind::Write(e) => {
                write!(f, "write error on shard {}: {}", self.shard_index, e)
            }
            StreamErrorKind::Codec(e) => {
                write!(f, "codec error on shard {}: {}", self.shard_index, e)
            }
        }
    }
}

impl std::error::Error for StreamError {
    fn source(&self) -> Option<&(dyn std::error::Error + 'static)> {
        match &self.kind {
            StreamErrorKind::Read(e) => Some(e),
            StreamErrorKind::Write(e) => Some(e),
            StreamErrorKind::Codec(e) => Some(e),
        }
    }
}

impl StreamError {
    fn read(shard_index: usize, e: std::io::Error) -> Self {
        Self {
            shard_index,
            kind: StreamErrorKind::Read(e),
        }
    }

    fn write(shard_index: usize, e: std::io::Error) -> Self {
        Self {
            shard_index,
            kind: StreamErrorKind::Write(e),
        }
    }

    fn codec(shard_index: usize, e: crate::Error) -> Self {
        Self {
            shard_index,
            kind: StreamErrorKind::Codec(e),
        }
    }
}

fn take_stream_error(
    first_error: &std::sync::Mutex<Option<StreamError>>,
    fallback: impl FnOnce() -> StreamError,
) -> StreamError {
    let mut guard = match first_error.lock() {
        Ok(guard) => guard,
        Err(poisoned) => poisoned.into_inner(),
    };
    guard.take().unwrap_or_else(fallback)
}

/// Record a parallel-write error, keeping the one with the smallest
/// `shard_index`. When multiple writers fail at once, the reported error no
/// longer depends on thread scheduling but deterministically points at the
/// lowest shard index, aiding diagnostics and reproduction.
fn record_min_error(first_error: &std::sync::Mutex<Option<StreamError>>, cand: StreamError) {
    if let Ok(mut guard) = first_error.lock() {
        let replace = match guard.as_ref() {
            Some(existing) => cand.shard_index < existing.shard_index,
            None => true,
        };
        if replace {
            *guard = Some(cand);
        }
    }
}

// ---------------------------------------------------------------------------
// Helpers
// ---------------------------------------------------------------------------

fn use_parallel_stream_io(options: &StreamOptions, stream_count: usize) -> bool {
    match options.io_mode {
        StreamIoMode::Serial => false,
        StreamIoMode::Parallel => true,
        StreamIoMode::Auto => use_parallel_stream_io_auto(options.block_size, stream_count),
    }
}

fn use_parallel_stream_io_auto(block_size: usize, stream_count: usize) -> bool {
    // Conservative single threshold: enable parallel I/O only when there are at
    // least 10 streams and blocks are at least 4 MiB. The two earlier
    // early-return guards (stream_count<2||block<256KiB and
    // stream_count<=6&&block<=1MiB) were strictly dominated by this predicate,
    // dead code that never changed the result, so they are removed to avoid
    // implying a tiered policy.
    block_size >= 4 * 1024 * 1024 && stream_count >= 10
}

fn read_block_with_mode<R: std::io::Read + Send>(
    readers: &mut [R],
    buffers: &mut [Vec<u8>],
    max_len: usize,
    use_parallel_io: bool,
    read_lengths: &mut Vec<(usize, usize)>,
) -> Result<(bool, usize), StreamError> {
    if use_parallel_io {
        read_block_par(readers, buffers, max_len)
    } else {
        read_block(readers, buffers, max_len, read_lengths)
    }
}

fn write_block_with_mode<W: std::io::Write + Send>(
    writers: &mut [W],
    buffers: &[Vec<u8>],
    len: usize,
    shard_offset: usize,
    use_parallel_io: bool,
) -> Result<(), StreamError> {
    if use_parallel_io {
        write_block_par(writers, buffers, len, shard_offset)
    } else {
        write_block(writers, buffers, len, shard_offset)
    }
}

/// Read up to `max_len` bytes from each reader into the corresponding
/// buffer, retrying on `Interrupted`.  Returns `Ok((all_eof, actual_len))`
/// where `all_eof` is `true` if every reader was already at EOF, and
/// `actual_len` is the number of bytes read (same across all readers,
/// with short reads zero-padded).
fn read_block<R: std::io::Read>(
    readers: &mut [R],
    buffers: &mut [Vec<u8>],
    max_len: usize,
    read_lengths: &mut Vec<(usize, usize)>,
) -> Result<(bool, usize), StreamError> {
    read_lengths.clear();

    for (i, (reader, buf)) in readers.iter_mut().zip(buffers.iter_mut()).enumerate() {
        let total = read_one_stream(reader, buf, max_len).map_err(|e| StreamError::read(i, e))?;
        read_lengths.push((i, total));
    }

    let actual_len = read_lengths
        .iter()
        .map(|(_, total)| *total)
        .max()
        .unwrap_or(0);
    zero_pad_short_buffers(buffers, read_lengths, actual_len);

    Ok((actual_len == 0, actual_len))
}

fn prepare_read_buffer(buf: &mut Vec<u8>, max_len: usize) {
    match buf.len().cmp(&max_len) {
        core::cmp::Ordering::Less => buf.resize(max_len, 0),
        core::cmp::Ordering::Greater => buf.truncate(max_len),
        core::cmp::Ordering::Equal => {}
    }
}

fn read_one_stream<R: std::io::Read>(
    reader: &mut R,
    buf: &mut Vec<u8>,
    max_len: usize,
) -> Result<usize, std::io::Error> {
    prepare_read_buffer(buf, max_len);
    let mut total = 0;

    while total < max_len {
        match reader.read(&mut buf[total..]) {
            Ok(0) => break,
            Ok(n) => total += n,
            Err(e) if e.kind() == std::io::ErrorKind::Interrupted => continue,
            Err(e) => return Err(e),
        }
    }

    Ok(total)
}

fn zero_pad_short_buffers(
    buffers: &mut [Vec<u8>],
    read_lengths: &[(usize, usize)],
    actual_len: usize,
) {
    for &(i, total) in read_lengths {
        if total < actual_len {
            buffers[i][total..actual_len].fill(0);
        }
    }
}

fn read_present_cursors_with_mode(
    shards: &mut [std::io::Cursor<Vec<u8>>],
    buffers: &mut [Vec<u8>],
    present: &[bool],
    present_indices: &[usize],
    max_len: usize,
    use_parallel_io: bool,
    read_lengths: &mut Vec<(usize, usize)>,
) -> Result<(bool, usize), StreamError> {
    read_lengths.clear();

    if use_parallel_io {
        use rayon::prelude::*;

        *read_lengths = shards
            .par_iter_mut()
            .zip(buffers.par_iter_mut())
            .enumerate()
            .filter_map(|(i, item)| present[i].then_some((i, item)))
            .map(|(i, (shard, buf))| {
                let total =
                    read_one_stream(shard, buf, max_len).map_err(|e| StreamError::read(i, e))?;
                buf.truncate(total);
                Ok::<_, StreamError>((i, total))
            })
            .collect::<Result<Vec<_>, _>>()?;
    } else {
        for &i in present_indices {
            let total = read_one_stream(&mut shards[i], &mut buffers[i], max_len)
                .map_err(|e| StreamError::read(i, e))?;
            buffers[i].truncate(total);
            read_lengths.push((i, total));
        }
    }

    let actual_len = read_lengths
        .iter()
        .map(|(_, total)| *total)
        .max()
        .unwrap_or(0);
    if actual_len != 0 {
        // Within a block, every present shard must read the same number of
        // bytes. A present shard that hits EOF early (fewer bytes than the
        // others) is truncated or length-mismatched and must not be silently
        // zero-padded into reconstruction, which would produce wrong recovered
        // data while returning Ok. This matches the in-memory `reconstruct`'s
        // `IncorrectShardSize` check.
        if let Some(&(bad, _)) = read_lengths.iter().find(|(_, total)| *total != actual_len) {
            return Err(StreamError::codec(bad, crate::Error::IncorrectShardSize));
        }
    }

    Ok((actual_len == 0, actual_len))
}

/// Write `len` bytes from each buffer to the corresponding writer.
fn write_block<W: std::io::Write>(
    writers: &mut [W],
    buffers: &[Vec<u8>],
    len: usize,
    shard_offset: usize,
) -> Result<(), StreamError> {
    for (i, (writer, buf)) in writers.iter_mut().zip(buffers.iter()).enumerate() {
        let mut written = 0;
        while written < len {
            match writer.write(&buf[written..len]) {
                Ok(0) => {
                    return Err(StreamError::write(
                        shard_offset + i,
                        std::io::Error::new(std::io::ErrorKind::WriteZero, "write returned 0"),
                    ));
                }
                Ok(n) => written += n,
                Err(e) if e.kind() == std::io::ErrorKind::Interrupted => continue,
                Err(e) => return Err(StreamError::write(shard_offset + i, e)),
            }
        }
    }
    Ok(())
}

// ---------------------------------------------------------------------------
// Parallel I/O helpers (rayon)
// ---------------------------------------------------------------------------

/// Parallel version of `read_block` — reads all shards concurrently.
fn read_block_par<R: std::io::Read + Send>(
    readers: &mut [R],
    buffers: &mut [Vec<u8>],
    max_len: usize,
) -> Result<(bool, usize), StreamError> {
    use rayon::prelude::*;

    let read_lengths: Vec<(usize, usize)> = readers
        .par_iter_mut()
        .zip(buffers.par_iter_mut())
        .enumerate()
        .map(|(i, (reader, buf))| {
            read_one_stream(reader, buf, max_len)
                .map(|total| (i, total))
                .map_err(|e| StreamError::read(i, e))
        })
        .collect::<Result<Vec<_>, _>>()?;

    let actual_len = read_lengths
        .iter()
        .map(|(_, total)| *total)
        .max()
        .unwrap_or(0);
    zero_pad_short_buffers(buffers, &read_lengths, actual_len);

    Ok((actual_len == 0, actual_len))
}

/// Parallel version of `write_block` — writes all shards concurrently.
fn write_block_par<W: std::io::Write + Send>(
    writers: &mut [W],
    buffers: &[Vec<u8>],
    len: usize,
    shard_offset: usize,
) -> Result<(), StreamError> {
    use rayon::prelude::*;

    let first_error: std::sync::Mutex<Option<StreamError>> = std::sync::Mutex::new(None);

    writers
        .par_iter_mut()
        .zip(buffers.par_iter())
        .enumerate()
        .try_for_each(|(i, (writer, buf))| {
            let mut written = 0;
            while written < len {
                match writer.write(&buf[written..len]) {
                    Ok(0) => {
                        record_min_error(
                            &first_error,
                            StreamError::write(
                                shard_offset + i,
                                std::io::Error::new(
                                    std::io::ErrorKind::WriteZero,
                                    "write returned 0",
                                ),
                            ),
                        );
                        return Err(());
                    }
                    Ok(n) => written += n,
                    Err(e) if e.kind() == std::io::ErrorKind::Interrupted => continue,
                    Err(e) => {
                        record_min_error(&first_error, StreamError::write(shard_offset + i, e));
                        return Err(());
                    }
                }
            }
            Ok(())
        })
        .map_err(|()| {
            // The fallback must also be a write error (not a fabricated read(0))
            // and point at this batch's starting shard index.
            take_stream_error(&first_error, || {
                StreamError::write(
                    shard_offset,
                    Error::other("parallel stream writer error was not reported"),
                )
            })
        })
}

// ---------------------------------------------------------------------------
// ReedSolomon streaming methods
// ---------------------------------------------------------------------------

impl super::ReedSolomon<crate::galois_8::Field> {
    /// Stream-encode data from readers into parity writers.
    ///
    /// Reads data shards in blocks of `options.block_size` bytes, encodes
    /// each block, and writes the resulting parity blocks.  This avoids
    /// loading the entire dataset into memory.
    ///
    /// All data streams should supply the same number of bytes. When they do
    /// not, shorter streams are zero-padded to the longest stream's length for
    /// each block, and no original-length metadata is persisted. The generated
    /// parity therefore only round-trips within this streaming API; interop
    /// with the in-memory API (which rejects unequal lengths with
    /// [`crate::Error::IncorrectShardSize`]) or with other Reed-Solomon
    /// implementations requires the caller to track the original shard lengths
    /// out of band.
    ///
    /// # Errors
    ///
    /// Returns [`StreamError`] on I/O failure or codec error.
    ///
    /// # Limitations
    ///
    /// Only the classic Reed-Solomon family is supported. Leopard-family codecs
    /// return [`crate::Error::UnsupportedCodecFamily`] (wrapped in a
    /// [`StreamError`]), because their FFT engines require whole-shard,
    /// 64-byte-aligned processing that the block-wise streaming path cannot
    /// satisfy.
    pub fn encode_stream(
        &self,
        data: &mut [impl std::io::Read + Send],
        parity: &mut [impl std::io::Write + Send],
        options: &StreamOptions,
    ) -> Result<(), StreamError> {
        // block_size is a public field and may bypass with_block_size's clamp
        // (e.g. a struct literal). Clamp it to the valid range here: 0 would be
        // raised to the lower bound (otherwise reads hit EOF immediately and
        // silently produce empty output), and a huge value is capped to the
        // upper bound (otherwise Vec::with_capacity may OOM).
        let block_size = options
            .block_size
            .clamp(MIN_BLOCK_SIZE_BYTES, MAX_BLOCK_SIZE_BYTES);
        // Leopard families (GF8/GF16, both built on ReedSolomon<galois_8::Field>)
        // dispatch to FFT codecs that require whole-shard, 64-byte-aligned
        // processing: a non-64-multiple final block fails with IncorrectShardSize,
        // so streaming is effectively unusable. Reject them explicitly, matching
        // the documented limitation.
        if self.is_leopard_gf8_family() || self.is_leopard_gf16_family() {
            return Err(StreamError::codec(0, crate::Error::UnsupportedCodecFamily));
        }

        let data_count = self.data_shard_count;
        let parity_count = self.parity_shard_count;
        let use_parallel_read = use_parallel_stream_io(options, data_count);
        let use_parallel_write = use_parallel_stream_io(options, parity_count);

        // Validate stream counts at runtime (debug_assert is removed in release
        // builds, which would otherwise lead to slice out-of-bounds panics or
        // silent partial encoding). Consistent with the crate's
        // check_piece_count! used on non-streaming paths.
        if data.len() != data_count {
            return Err(StreamError::codec(0, crate::Error::TooFewShards));
        }
        if parity.len() != parity_count {
            return Err(StreamError::codec(0, crate::Error::TooFewShards));
        }

        let mut data_bufs: Vec<Vec<u8>> = (0..data_count)
            .map(|_| Vec::with_capacity(block_size))
            .collect();
        let mut parity_bufs: Vec<Vec<u8>> = (0..parity_count)
            .map(|_| Vec::with_capacity(block_size))
            .collect();
        let mut read_lengths = Vec::with_capacity(data_count);

        loop {
            let (all_eof, actual_len) = read_block_with_mode(
                data,
                &mut data_bufs,
                block_size,
                use_parallel_read,
                &mut read_lengths,
            )?;
            if all_eof {
                break;
            }

            // Resize parity buffers to match actual length.
            for buf in parity_bufs.iter_mut() {
                buf.resize(actual_len, 0);
            }

            // Encode (parallel codec).
            let data_refs: Vec<&[u8]> = data_bufs.iter().map(|b| &b[..actual_len]).collect();
            let mut parity_refs: Vec<&mut [u8]> = parity_bufs
                .iter_mut()
                .map(|b| &mut b[..actual_len])
                .collect();

            self.encode_sep_par(&data_refs, &mut parity_refs)
                .map_err(|e| StreamError::codec(0, e))?;

            // Write parity using the selected stream I/O mode.
            write_block_with_mode(
                parity,
                &parity_bufs,
                actual_len,
                data_count,
                use_parallel_write,
            )?;
        }

        Ok(())
    }

    /// Stream-verify data + parity from readers.
    ///
    /// Reads all shards in blocks, verifying each block independently.
    /// Returns `Ok(true)` if every block is valid, `Ok(false)` if any block
    /// fails verification.
    ///
    /// # Limitations
    ///
    /// Only the classic Reed-Solomon family is supported; Leopard-family codecs
    /// return [`crate::Error::UnsupportedCodecFamily`].
    pub fn verify_stream(
        &self,
        shards: &mut [impl std::io::Read + Send],
        options: &StreamOptions,
    ) -> Result<bool, StreamError> {
        // block_size is a public field and may bypass with_block_size's clamp
        // (e.g. a struct literal). Clamp it to the valid range here: 0 would be
        // raised to the lower bound (otherwise reads hit EOF immediately and
        // silently produce empty output), and a huge value is capped to the
        // upper bound (otherwise Vec::with_capacity may OOM).
        let block_size = options
            .block_size
            .clamp(MIN_BLOCK_SIZE_BYTES, MAX_BLOCK_SIZE_BYTES);
        // Leopard families are unsupported for streaming (see encode_stream).
        if self.is_leopard_gf8_family() || self.is_leopard_gf16_family() {
            return Err(StreamError::codec(0, crate::Error::UnsupportedCodecFamily));
        }

        let total = self.total_shard_count;
        let use_parallel_read = use_parallel_stream_io(options, total);

        // Validate the shard count at runtime (debug_assert is removed in
        // release builds).
        if shards.len() != total {
            return Err(StreamError::codec(0, crate::Error::TooFewShards));
        }

        let mut bufs: Vec<Vec<u8>> = (0..total).map(|_| Vec::with_capacity(block_size)).collect();
        let mut read_lengths = Vec::with_capacity(total);

        loop {
            let (all_eof, actual_len) = read_block_with_mode(
                shards,
                &mut bufs,
                block_size,
                use_parallel_read,
                &mut read_lengths,
            )?;
            if all_eof {
                break;
            }

            let refs: Vec<&[u8]> = bufs.iter().map(|b| &b[..actual_len]).collect();

            let valid = self
                .verify_par(&refs)
                .map_err(|e| StreamError::codec(0, e))?;
            if !valid {
                return Ok(false);
            }
        }

        Ok(true)
    }

    /// Stream-reconstruct missing shards.
    ///
    /// `shards` has one entry per total shard.  Present shards contain their
    /// data in a `Cursor<Vec<u8>>`; missing shards use an empty cursor
    /// (`Cursor::new(Vec::new())`).  Recovered data is written into the
    /// missing shards' cursors.
    ///
    /// Present cursors are read from the beginning: their position is reset to
    /// `0` before reading, so it is safe to pass cursors that were just written
    /// to (whose position sits at the end).
    ///
    /// The function reads blocks from present shards, reconstructs missing
    /// blocks, and writes recovered data into the missing cursors.  The set
    /// of missing shard indices must be consistent across all blocks.
    ///
    /// # Limitations
    ///
    /// Only the classic Reed-Solomon family is supported; Leopard-family codecs
    /// return [`crate::Error::UnsupportedCodecFamily`] rather than silently
    /// producing incorrect results.
    pub fn reconstruct_stream(
        &self,
        shards: &mut [std::io::Cursor<Vec<u8>>],
        options: &StreamOptions,
    ) -> Result<(), StreamError> {
        // block_size is a public field and may bypass with_block_size's clamp
        // (e.g. a struct literal). Clamp it to the valid range here: 0 would be
        // raised to the lower bound (otherwise reads hit EOF immediately and
        // silently produce empty output), and a huge value is capped to the
        // upper bound (otherwise Vec::with_capacity may OOM).
        let block_size = options
            .block_size
            .clamp(MIN_BLOCK_SIZE_BYTES, MAX_BLOCK_SIZE_BYTES);
        let total = self.total_shard_count;

        // Validate the shard count at runtime (debug_assert is removed in
        // release builds).
        if shards.len() != total {
            return Err(StreamError::codec(0, crate::Error::TooFewShards));
        }

        // Leopard families are unsupported for streaming (see encode_stream).
        if self.is_leopard_gf8_family() || self.is_leopard_gf16_family() {
            return Err(StreamError::codec(0, crate::Error::UnsupportedCodecFamily));
        }

        // Determine which shards are present (non-empty cursor).
        let mut present = vec![false; total];
        for (i, shard) in shards.iter().enumerate() {
            present[i] = !shard.get_ref().is_empty();
        }

        let missing_count = present.iter().filter(|&&p| !p).count();
        let present_count = total - missing_count;
        // An empty dataset (no present shards) has nothing to recover; return Ok,
        // consistent with encode_stream returning Ok for empty input (rather than
        // TooFewShardsPresent).
        if present_count == 0 {
            return Ok(());
        }
        if missing_count > self.parity_shard_count {
            return Err(StreamError::codec(0, crate::Error::TooFewShardsPresent));
        }
        let use_parallel_read = use_parallel_stream_io(options, present_count);
        let present_indices: Vec<usize> = present
            .iter()
            .enumerate()
            .filter_map(|(i, is_present)| is_present.then_some(i))
            .collect();
        let missing_indices: Vec<usize> = present
            .iter()
            .enumerate()
            .filter_map(|(i, is_present)| (!is_present).then_some(i))
            .collect();

        // Presence is decided from whether the underlying Vec is empty, but
        // reads go through the Cursor's Read impl (from the current position).
        // A present cursor whose position is not 0 (e.g. just written to, with
        // the position at the end) would be misread as empty, causing silent
        // no-recovery or wrong recovery. Reset every present cursor's position
        // to 0 before reading.
        for &idx in &present_indices {
            shards[idx].set_position(0);
        }

        // Strategy: read present shards into buffers per block, reconstruct,
        // then write recovered data into missing shards' cursors.

        // Pre-allocate read buffers and the reconstruct container once. Present
        // shard buffers are moved into the reconstruct call and then moved back
        // after each block so their allocations survive across iterations.
        let mut bufs: Vec<Vec<u8>> = (0..total)
            .map(|i| {
                if present[i] {
                    Vec::with_capacity(block_size)
                } else {
                    Vec::new()
                }
            })
            .collect();
        let mut reconstruct_bufs: Vec<Option<Vec<u8>>> = (0..total).map(|_| None).collect();
        let mut read_lengths = Vec::with_capacity(total);

        loop {
            let (all_eof, actual_len) = read_present_cursors_with_mode(
                shards,
                &mut bufs,
                &present,
                &present_indices,
                block_size,
                use_parallel_read,
                &mut read_lengths,
            )?;
            if all_eof {
                break;
            }

            // Zero-pad present shards to actual_len and reuse the outer
            // Option<Vec<_>> container for reconstructing this block.
            for &idx in &present_indices {
                bufs[idx].resize(actual_len, 0);
                reconstruct_bufs[idx] = Some(mem::take(&mut bufs[idx]));
            }
            for &idx in &missing_indices {
                reconstruct_bufs[idx] = None;
            }

            self.reconstruct(&mut reconstruct_bufs)
                .map_err(|e| StreamError::codec(0, e))?;

            // Write recovered data into missing shards' cursors.
            // (reconstruct fills in all missing shards — data and parity)
            for &idx in &missing_indices {
                let recovered = reconstruct_bufs[idx]
                    .as_ref()
                    .expect("missing shard buffer");
                shards[idx]
                    .get_mut()
                    .extend_from_slice(&recovered[..actual_len]);
            }

            for idx in 0..total {
                if let Some(buf) = reconstruct_bufs[idx].take() {
                    bufs[idx] = buf;
                }
            }
        }

        Ok(())
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::galois_8::ReedSolomon;

    fn make_codec(data: usize, parity: usize) -> ReedSolomon {
        ReedSolomon::new(data, parity).unwrap()
    }

    fn random_data(len: usize) -> Vec<u8> {
        // Simple deterministic fill for reproducibility.
        (0..len)
            .map(|i| (i.wrapping_mul(73).wrapping_add(17)) as u8)
            .collect()
    }

    #[test]
    fn test_encode_stream_basic() {
        let rs = make_codec(3, 2);
        let shard_size = 4096;

        let d0 = random_data(shard_size);
        let d1 = random_data(shard_size);
        let d2 = random_data(shard_size);

        let mut readers: Vec<&[u8]> = vec![d0.as_slice(), d1.as_slice(), d2.as_slice()];
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(), Vec::new()];

        rs.encode_stream(&mut readers, &mut writers, &StreamOptions::default())
            .unwrap();

        assert_eq!(writers[0].len(), shard_size);
        assert_eq!(writers[1].len(), shard_size);

        // Verify parity is correct by verifying the full shard set.
        let all: Vec<&[u8]> = vec![
            d0.as_slice(),
            d1.as_slice(),
            d2.as_slice(),
            writers[0].as_slice(),
            writers[1].as_slice(),
        ];
        assert!(rs.verify(&all).unwrap());
    }

    #[test]
    fn test_encode_stream_multi_block() {
        let rs = make_codec(3, 2);
        let total_size = 10 * 1024; // 10 KiB
        let block_size = 4096;

        let d0 = random_data(total_size);
        let d1 = random_data(total_size);
        let d2 = random_data(total_size);

        let mut readers: Vec<&[u8]> = vec![d0.as_slice(), d1.as_slice(), d2.as_slice()];
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(), Vec::new()];

        let opts = StreamOptions::new().with_block_size(block_size);
        rs.encode_stream(&mut readers, &mut writers, &opts).unwrap();

        assert_eq!(writers[0].len(), total_size);
        assert_eq!(writers[1].len(), total_size);
    }

    #[test]
    fn test_encode_stream_empty() {
        let rs = make_codec(3, 2);
        let empty: Vec<&[u8]> = vec![&[], &[], &[]];
        let mut readers = empty.clone();
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(), Vec::new()];

        rs.encode_stream(&mut readers, &mut writers, &StreamOptions::default())
            .unwrap();

        // Empty input → empty output.
        assert!(writers[0].is_empty());
        assert!(writers[1].is_empty());
    }

    #[test]
    fn test_encode_stream_unequal_lengths() {
        let rs = make_codec(3, 2);

        let d0 = random_data(1000);
        let d1 = random_data(500);
        let d2 = random_data(800);

        let mut readers: Vec<&[u8]> = vec![d0.as_slice(), d1.as_slice(), d2.as_slice()];
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(), Vec::new()];

        rs.encode_stream(&mut readers, &mut writers, &StreamOptions::default())
            .unwrap();

        // All outputs should be 1000 bytes (max of input lengths, since we
        // zero-pad short shards).
        assert_eq!(writers[0].len(), 1000);
        assert_eq!(writers[1].len(), 1000);
    }

    #[test]
    fn test_encode_stream_10x4() {
        let rs = make_codec(10, 4);
        let shard_size = 64 * 1024;

        let data: Vec<Vec<u8>> = (0..10).map(|_| random_data(shard_size)).collect();
        let mut readers: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(); 4];

        rs.encode_stream(&mut readers, &mut writers, &StreamOptions::default())
            .unwrap();

        // Verify.
        let mut all: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
        for w in &writers {
            all.push(w.as_slice());
        }
        assert!(rs.verify(&all).unwrap());
    }

    #[test]
    fn test_verify_stream_valid() {
        let rs = make_codec(3, 2);
        let shard_size = 4096;

        let d = vec![random_data(shard_size); 3];
        let mut readers: Vec<&[u8]> = d.iter().map(|s| s.as_slice()).collect();
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(); 2];
        rs.encode_stream(&mut readers, &mut writers, &StreamOptions::default())
            .unwrap();

        // Build full shard set for verification.
        let mut all_data: Vec<&[u8]> = d.iter().map(|s| s.as_slice()).collect();
        for w in &writers {
            all_data.push(w.as_slice());
        }
        let mut all_readers: Vec<&[u8]> = all_data;

        assert!(
            rs.verify_stream(&mut all_readers, &StreamOptions::default())
                .unwrap()
        );
    }

    #[test]
    fn test_verify_stream_corrupted() {
        let rs = make_codec(3, 2);
        let shard_size = 4096;

        let d = vec![random_data(shard_size); 3];
        let mut readers: Vec<&[u8]> = d.iter().map(|s| s.as_slice()).collect();
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(); 2];
        rs.encode_stream(&mut readers, &mut writers, &StreamOptions::default())
            .unwrap();

        // Corrupt a parity shard.
        writers[0][0] ^= 0xFF;

        let mut all_data: Vec<&[u8]> = d.iter().map(|s| s.as_slice()).collect();
        for w in &writers {
            all_data.push(w.as_slice());
        }
        let mut all_readers: Vec<&[u8]> = all_data;

        assert!(
            !rs.verify_stream(&mut all_readers, &StreamOptions::default())
                .unwrap()
        );
    }

    #[test]
    fn test_reconstruct_stream_single_missing() {
        let rs = make_codec(3, 2);
        let shard_size = 4096;

        let d: Vec<Vec<u8>> = (0..3).map(|_| random_data(shard_size)).collect();
        let mut readers: Vec<&[u8]> = d.iter().map(|s| s.as_slice()).collect();
        let mut parity_writers: Vec<Vec<u8>> = vec![Vec::new(); 2];
        rs.encode_stream(&mut readers, &mut parity_writers, &StreamOptions::default())
            .unwrap();

        // data[0] is missing — use empty Cursor.
        let mut shards: Vec<std::io::Cursor<Vec<u8>>> = vec![
            std::io::Cursor::new(Vec::new()), // missing
            std::io::Cursor::new(d[1].clone()),
            std::io::Cursor::new(d[2].clone()),
            std::io::Cursor::new(parity_writers[0].clone()),
            std::io::Cursor::new(parity_writers[1].clone()),
        ];

        rs.reconstruct_stream(&mut shards, &StreamOptions::default())
            .unwrap();

        // Shard 0 should have been recovered into the cursor's inner Vec.
        assert_eq!(shards[0].get_ref(), &d[0]);
    }

    #[test]
    fn test_reconstruct_non_streaming() {
        // Verify basic encode + reconstruct works without streaming.
        let rs = make_codec(3, 2);
        let shard_size = 4096;

        let d: Vec<Vec<u8>> = (0..3).map(|_| random_data(shard_size)).collect();
        let data_refs: Vec<&[u8]> = d.iter().map(|s| s.as_slice()).collect();
        let mut p0 = vec![0u8; shard_size];
        let mut p1 = vec![0u8; shard_size];
        let mut parity_refs: Vec<&mut [u8]> = vec![&mut p0, &mut p1];
        rs.encode_sep(&data_refs, &mut parity_refs).unwrap();

        // Now reconstruct with shard 0 missing.
        let mut shards: Vec<Option<Vec<u8>>> = vec![
            None,
            Some(d[1].clone()),
            Some(d[2].clone()),
            Some(p0.clone()),
            Some(p1.clone()),
        ];
        rs.reconstruct(&mut shards).unwrap();

        assert_eq!(shards[0].as_ref().unwrap(), &d[0]);
    }

    #[test]
    fn test_reconstruct_stream_basic() {
        let rs = make_codec(3, 2);
        let shard_size = 64;

        let d: Vec<Vec<u8>> = (0..3).map(|_| random_data(shard_size)).collect();
        let mut readers: Vec<&[u8]> = d.iter().map(|s| s.as_slice()).collect();
        let mut parity_writers: Vec<Vec<u8>> = vec![Vec::new(); 2];
        rs.encode_stream(&mut readers, &mut parity_writers, &StreamOptions::default())
            .unwrap();

        // Verify parity is correct.
        let all: Vec<&[u8]> = vec![
            d[0].as_slice(),
            d[1].as_slice(),
            d[2].as_slice(),
            parity_writers[0].as_slice(),
            parity_writers[1].as_slice(),
        ];
        assert!(rs.verify(&all).unwrap());

        // Missing shard 0 — empty Cursor; present shards have data.
        let mut shards: Vec<std::io::Cursor<Vec<u8>>> = vec![
            std::io::Cursor::new(Vec::new()), // missing
            std::io::Cursor::new(d[1].clone()),
            std::io::Cursor::new(d[2].clone()),
            std::io::Cursor::new(parity_writers[0].clone()),
            std::io::Cursor::new(parity_writers[1].clone()),
        ];

        rs.reconstruct_stream(&mut shards, &StreamOptions::default())
            .unwrap();

        // Shard 0 recovered into the empty cursor's inner Vec.
        let recovered = shards[0].get_ref();
        assert_eq!(
            recovered.len(),
            d[0].len(),
            "recovered len {} != expected len {}",
            recovered.len(),
            d[0].len()
        );
        assert_eq!(
            recovered,
            &d[0],
            "recovered: {:?}, expected: {:?}",
            &recovered[..8],
            &d[0][..8]
        );
    }

    #[test]
    fn test_stream_options_builder() {
        let opts = StreamOptions::new().with_block_size(1024 * 1024);
        assert_eq!(opts.block_size, 1024 * 1024);
        assert_eq!(opts.io_mode, StreamIoMode::Auto);

        // Minimum is 1 KiB.
        let opts = StreamOptions::new().with_block_size(100);
        assert_eq!(opts.block_size, 1024);

        let opts = StreamOptions::new().with_io_mode(StreamIoMode::Serial);
        assert_eq!(opts.io_mode, StreamIoMode::Serial);
    }

    #[test]
    fn test_stream_io_auto_decision_thresholds() {
        let opts = StreamOptions::new()
            .with_block_size(64 * 1024)
            .with_io_mode(StreamIoMode::Auto);
        assert!(!use_parallel_stream_io(&opts, 14));

        let opts = StreamOptions::new()
            .with_block_size(1024 * 1024)
            .with_io_mode(StreamIoMode::Auto);
        assert!(!use_parallel_stream_io(&opts, 6));

        let opts = StreamOptions::new()
            .with_block_size(4 * 1024 * 1024)
            .with_io_mode(StreamIoMode::Auto);
        assert!(use_parallel_stream_io(&opts, 10));

        let opts = StreamOptions::new()
            .with_block_size(64 * 1024)
            .with_io_mode(StreamIoMode::Parallel);
        assert!(use_parallel_stream_io(&opts, 1));
    }

    #[test]
    fn test_stream_error_display() {
        let e = StreamError::read(
            3,
            std::io::Error::new(std::io::ErrorKind::UnexpectedEof, "eof"),
        );
        let s = format!("{e}");
        assert!(s.contains("shard 3"));
        assert!(s.contains("eof"));

        let e = StreamError::codec(0, crate::Error::TooFewShardsPresent);
        let s = format!("{e}");
        assert!(s.contains("codec"));
    }

    // -----------------------------------------------------------------------
    // Concurrent stream tests (P0-2e-3)
    // -----------------------------------------------------------------------

    #[test]
    fn test_encode_stream_concurrent() {
        let rs = make_codec(4, 2);
        let shard_size = 8 * 1024;

        let data: Vec<Vec<u8>> = (0..4).map(|_| random_data(shard_size)).collect();
        let mut readers: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(); 2];

        rs.encode_stream(&mut readers, &mut writers, &StreamOptions::default())
            .unwrap();

        // Verify parity correctness.
        let mut all: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
        for w in &writers {
            all.push(w.as_slice());
        }
        assert!(rs.verify(&all).unwrap());
    }

    #[test]
    fn test_verify_stream_concurrent() {
        let rs = make_codec(4, 2);
        let shard_size = 8 * 1024;

        let data: Vec<Vec<u8>> = (0..4).map(|_| random_data(shard_size)).collect();
        let mut readers: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(); 2];
        rs.encode_stream(&mut readers, &mut writers, &StreamOptions::default())
            .unwrap();

        // Valid case.
        let mut all: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
        for w in &writers {
            all.push(w.as_slice());
        }
        let mut all_readers: Vec<&[u8]> = all;
        assert!(
            rs.verify_stream(&mut all_readers, &StreamOptions::default())
                .unwrap()
        );

        // Corrupted case.
        let mut corrupted = writers[0].clone();
        corrupted[0] ^= 0xFF;
        let mut all_corrupt: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
        all_corrupt.push(corrupted.as_slice());
        all_corrupt.push(writers[1].as_slice());
        assert!(
            !rs.verify_stream(&mut all_corrupt, &StreamOptions::default())
                .unwrap()
        );
    }

    #[test]
    fn test_reconstruct_stream_concurrent() {
        let rs = make_codec(4, 2);
        let shard_size = 8 * 1024;

        let data: Vec<Vec<u8>> = (0..4).map(|_| random_data(shard_size)).collect();
        let mut readers: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
        let mut parity_writers: Vec<Vec<u8>> = vec![Vec::new(); 2];
        rs.encode_stream(&mut readers, &mut parity_writers, &StreamOptions::default())
            .unwrap();

        // Missing shards: data[1] and parity[0].
        let mut shards: Vec<std::io::Cursor<Vec<u8>>> = vec![
            std::io::Cursor::new(data[0].clone()),
            std::io::Cursor::new(Vec::new()), // missing
            std::io::Cursor::new(data[2].clone()),
            std::io::Cursor::new(data[3].clone()),
            std::io::Cursor::new(Vec::new()), // missing
            std::io::Cursor::new(parity_writers[1].clone()),
        ];

        rs.reconstruct_stream(&mut shards, &StreamOptions::default())
            .unwrap();

        assert_eq!(shards[1].get_ref(), &data[1]);
    }

    #[test]
    fn test_concurrent_stream_large_blocks() {
        let rs = make_codec(10, 4);
        let total_size = 1024 * 1024; // 1 MiB
        let block_size = 256 * 1024; // 256 KiB blocks

        let data: Vec<Vec<u8>> = (0..10).map(|_| random_data(total_size)).collect();
        let mut readers: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(); 4];

        let opts = StreamOptions::new().with_block_size(block_size);
        rs.encode_stream(&mut readers, &mut writers, &opts).unwrap();

        // Verify all blocks.
        let mut all: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
        for w in &writers {
            all.push(w.as_slice());
        }
        assert!(rs.verify(&all).unwrap());

        // Reconstruct with 2 missing data shards.
        let mut shards: Vec<std::io::Cursor<Vec<u8>>> = Vec::new();
        for d in &data {
            shards.push(std::io::Cursor::new(d.clone()));
        }
        shards[0] = std::io::Cursor::new(Vec::new());
        shards[5] = std::io::Cursor::new(Vec::new());
        for w in &writers {
            shards.push(std::io::Cursor::new(w.clone()));
        }

        rs.reconstruct_stream(&mut shards, &StreamOptions::default())
            .unwrap();

        assert_eq!(shards[0].get_ref(), &data[0]);
        assert_eq!(shards[5].get_ref(), &data[5]);
    }

    #[test]
    fn test_encode_stream_zero_length() {
        let rs = ReedSolomon::new(3, 2).unwrap();
        let mut readers: Vec<&[u8]> = vec![b""; 3];
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(); 2];
        // Zero-length data should succeed (encode produces empty parity).
        rs.encode_stream(&mut readers, &mut writers, &StreamOptions::default())
            .unwrap();
        assert!(writers.iter().all(|w| w.is_empty()));
    }

    #[test]
    fn test_encode_stream_single_byte_block() {
        let rs = ReedSolomon::new(2, 1).unwrap();
        let d0 = vec![0xABu8; 4];
        let d1 = vec![0xCDu8; 4];
        let mut readers: Vec<&[u8]> = vec![d0.as_slice(), d1.as_slice()];
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(); 1];
        let opts = StreamOptions::new().with_block_size(1);
        rs.encode_stream(&mut readers, &mut writers, &opts).unwrap();

        // Verify with the smallest possible block size.
        let mut all: Vec<&[u8]> = vec![d0.as_slice(), d1.as_slice()];
        for w in &writers {
            all.push(w.as_slice());
        }
        assert!(rs.verify(&all).unwrap());
    }

    #[test]
    fn test_stream_io_modes_encode_verify_match() {
        let rs = ReedSolomon::new(4, 2).unwrap();
        let shard_size = 32 * 1024;
        let data: Vec<Vec<u8>> = (0..4).map(|_| random_data(shard_size)).collect();

        let mut expected_parity = Vec::new();
        for mode in [
            StreamIoMode::Auto,
            StreamIoMode::Serial,
            StreamIoMode::Parallel,
        ] {
            let opts = StreamOptions::new()
                .with_block_size(8 * 1024)
                .with_io_mode(mode);
            let mut readers: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
            let mut writers: Vec<Vec<u8>> = vec![Vec::new(); 2];

            rs.encode_stream(&mut readers, &mut writers, &opts).unwrap();
            if expected_parity.is_empty() {
                expected_parity = writers.clone();
            } else {
                assert_eq!(writers, expected_parity);
            }

            let mut all: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
            for parity in &writers {
                all.push(parity.as_slice());
            }
            let mut verify_readers = all;
            assert!(rs.verify_stream(&mut verify_readers, &opts).unwrap());
        }
    }

    #[test]
    fn test_reconstruct_stream_minimum_present() {
        let rs = ReedSolomon::new(3, 2).unwrap();
        let shard_len = 16usize;
        let data: Vec<Vec<u8>> = (0..3).map(|i| vec![i as u8 + 1; shard_len]).collect();

        // Encode to get parity.
        let refs: Vec<&[u8]> = data.iter().map(|d| d.as_slice()).collect();
        let mut parity = vec![vec![0u8; shard_len]; 2];
        let mut par_refs: Vec<&mut [u8]> = parity.iter_mut().map(|p| &mut p[..]).collect();
        rs.encode_sep(&refs, &mut par_refs).unwrap();

        // Keep minimum viable: 3 of 5 shards (data_shard_count).
        // Present: shard 0 (data), shard 3 (parity), shard 4 (parity).
        // Missing: shard 1 (data), shard 2 (data).
        let mut shards: Vec<std::io::Cursor<Vec<u8>>> = vec![
            std::io::Cursor::new(data[0].clone()),
            std::io::Cursor::new(Vec::new()),
            std::io::Cursor::new(Vec::new()),
            std::io::Cursor::new(parity[0].clone()),
            std::io::Cursor::new(parity[1].clone()),
        ];

        rs.reconstruct_stream(&mut shards, &StreamOptions::default())
            .unwrap();

        // Verify reconstructed data shards match originals.
        assert_eq!(shards[1].get_ref(), &data[1], "shard 1 mismatch");
        assert_eq!(shards[2].get_ref(), &data[2], "shard 2 mismatch");
    }

    // -----------------------------------------------------------------------
    // Streaming-path audit regression tests
    // -----------------------------------------------------------------------

    fn leopard_gf8_codec(data: usize, parity: usize) -> ReedSolomon {
        use crate::{CodecFamily, CodecOptions};
        ReedSolomon::with_options(
            data,
            parity,
            CodecOptions {
                codec_family: CodecFamily::LeopardGF8,
                ..CodecOptions::default()
            },
        )
        .unwrap()
    }

    /// A truncated / length-mismatched present shard must error with
    /// `IncorrectShardSize` instead of silently recovering wrong data.
    #[test]
    fn test_reconstruct_stream_truncated_present_errors() {
        let rs = make_codec(3, 2);
        let shard = 4096;
        let d: Vec<Vec<u8>> = (0..3).map(|_| random_data(shard)).collect();
        let mut readers: Vec<&[u8]> = d.iter().map(|s| s.as_slice()).collect();
        let mut parity: Vec<Vec<u8>> = vec![Vec::new(); 2];
        rs.encode_stream(&mut readers, &mut parity, &StreamOptions::default())
            .unwrap();

        // shard0 missing; shard1 truncated to 50 bytes.
        let mut shards: Vec<std::io::Cursor<Vec<u8>>> = vec![
            std::io::Cursor::new(Vec::new()),
            std::io::Cursor::new(d[1][..50].to_vec()),
            std::io::Cursor::new(d[2].clone()),
            std::io::Cursor::new(parity[0].clone()),
            std::io::Cursor::new(parity[1].clone()),
        ];
        let err = rs
            .reconstruct_stream(&mut shards, &StreamOptions::default())
            .unwrap_err();
        assert!(matches!(
            err.kind,
            StreamErrorKind::Codec(crate::Error::IncorrectShardSize)
        ));
    }

    /// A present cursor whose position is not 0 (written first) still recovers
    /// correctly.
    #[test]
    fn test_reconstruct_stream_nonzero_cursor_position() {
        use std::io::Write;
        let rs = make_codec(3, 2);
        let shard = 4096;
        let d: Vec<Vec<u8>> = (0..3).map(|_| random_data(shard)).collect();
        let mut readers: Vec<&[u8]> = d.iter().map(|s| s.as_slice()).collect();
        let mut parity: Vec<Vec<u8>> = vec![Vec::new(); 2];
        rs.encode_stream(&mut readers, &mut parity, &StreamOptions::default())
            .unwrap();

        let mk = |bytes: &[u8]| {
            let mut c = std::io::Cursor::new(Vec::new());
            c.write_all(bytes).unwrap(); // position left at the end
            c
        };
        let mut shards = vec![
            std::io::Cursor::new(Vec::new()), // missing
            mk(&d[1]),
            mk(&d[2]),
            mk(&parity[0]),
            mk(&parity[1]),
        ];
        rs.reconstruct_stream(&mut shards, &StreamOptions::default())
            .unwrap();
        assert_eq!(shards[0].get_ref(), &d[0]);
    }

    /// `block_size = 0` (public field) is clamped and encodes normally instead
    /// of silently producing empty output.
    #[test]
    fn test_encode_stream_zero_block_size_clamped() {
        let rs = make_codec(3, 2);
        let shard = 4096;
        let d: Vec<Vec<u8>> = (0..3).map(|_| random_data(shard)).collect();
        let mut readers: Vec<&[u8]> = d.iter().map(|s| s.as_slice()).collect();
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(); 2];
        let opts = StreamOptions {
            block_size: 0,
            io_mode: StreamIoMode::Serial,
        };
        rs.encode_stream(&mut readers, &mut writers, &opts).unwrap();
        assert_eq!(writers[0].len(), shard);
        assert_eq!(writers[1].len(), shard);
    }

    /// A mismatched stream count returns an error at runtime (no panic, no
    /// reliance on `debug_assert`).
    #[test]
    fn test_encode_stream_wrong_count_errors() {
        let rs = make_codec(3, 2);
        let d0 = random_data(64);
        let mut readers: Vec<&[u8]> = vec![d0.as_slice(), d0.as_slice()]; // one short
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(); 2];
        let err = rs
            .encode_stream(&mut readers, &mut writers, &StreamOptions::default())
            .unwrap_err();
        assert!(matches!(
            err.kind,
            StreamErrorKind::Codec(crate::Error::TooFewShards)
        ));
    }

    /// Leopard families are rejected by all three streaming methods with
    /// `UnsupportedCodecFamily`.
    #[test]
    fn test_stream_rejects_leopard_family() {
        let rs = leopard_gf8_codec(4, 2);
        let d: Vec<Vec<u8>> = (0..6).map(|_| random_data(128)).collect();

        let mut enc_readers: Vec<&[u8]> = d[..4].iter().map(|s| s.as_slice()).collect();
        let mut writers: Vec<Vec<u8>> = vec![Vec::new(); 2];
        let err = rs
            .encode_stream(&mut enc_readers, &mut writers, &StreamOptions::default())
            .unwrap_err();
        assert!(matches!(
            err.kind,
            StreamErrorKind::Codec(crate::Error::UnsupportedCodecFamily)
        ));

        let mut ver_readers: Vec<&[u8]> = d.iter().map(|s| s.as_slice()).collect();
        let err = rs
            .verify_stream(&mut ver_readers, &StreamOptions::default())
            .unwrap_err();
        assert!(matches!(
            err.kind,
            StreamErrorKind::Codec(crate::Error::UnsupportedCodecFamily)
        ));

        let mut rec_shards: Vec<std::io::Cursor<Vec<u8>>> =
            d.iter().map(|s| std::io::Cursor::new(s.clone())).collect();
        let err = rs
            .reconstruct_stream(&mut rec_shards, &StreamOptions::default())
            .unwrap_err();
        assert!(matches!(
            err.kind,
            StreamErrorKind::Codec(crate::Error::UnsupportedCodecFamily)
        ));
    }

    /// An empty dataset (all empty cursors) reconstructs to Ok, consistent with
    /// `encode_stream` on empty input.
    #[test]
    fn test_reconstruct_stream_empty_dataset_ok() {
        let rs = make_codec(3, 2);
        let mut shards: Vec<std::io::Cursor<Vec<u8>>> =
            (0..5).map(|_| std::io::Cursor::new(Vec::new())).collect();
        rs.reconstruct_stream(&mut shards, &StreamOptions::default())
            .unwrap();
        assert!(shards.iter().all(|c| c.get_ref().is_empty()));
    }
}