dar_forensic/

lib.rs

//! Pure-Rust reader for Denis Corbin DAR (Disk ARchiver) archives.
//!
//! Supports DAR formats 7–11 (produced by dar 2.3–2.8) and the legacy ≤7 grammar.
//! Passware Kit Mobile produces format-9 archives; dar 2.8.5 produces 11.3.
//! Entries and the catalogue compressed with gzip, bzip2 or xz are transparently
//! decompressed (pure-Rust); lzo, zstd, lz4 and encryption are not decoded.
//!
//! ## Format sketch
//!
//! ```text
//! Slice header:
//!   [4]  magic = 00 00 00 7b  (SAUV_MAGIC_NUMBER = 123, big-endian u32)
//!   [10] internal_name label
//!   [1]  flag  [1]  ext_char
//!   TLV list:  infinint(count) + count × (u16 type + infinint len + data)
//!   ← archive_origin: all catalog archive_offset values are relative to here
//!
//! Archive body:
//!   escaped sequences (seqt_file, seqt_saved, …) + raw file bytes
//!
//! Catalog  (located by seqt_catalogue escape: AD FD EA 77 21 43):
//!   [10] label  +  (NUL working-dir path, format 11.1+ only)  +  entries
//!
//!   Each entry: cat_sig byte where (cat_sig & 0x1f | 0x60) gives type
//!     'd' directory  → NUL-name + inode [+ FSA]  (push to dir stack)
//!     'f' file       → NUL-name + inode [+ FSA] + file-specific fields
//!     'z' EOD        → pop dir stack; depth=0 → done
//! ```
//!
//! ## Key non-obvious invariants
//!
//! - **Infinint**: variable-length. The common form is 5 bytes
//!   (`0x80 XX XX XX XX`, a big-endian u32); timestamps past 2^32 use the
//!   9-byte `0x40` form (big-endian u64). Encodings wider than 64 bits are
//!   rejected as corrupt — this reader decodes to `u64` or errors, never
//!   truncates.
//! - **Permissions**: 2-byte big-endian u16, *not* an infinint.
//! - **Timestamps**: format 8 stores a bare seconds infinint; format 9+ prefix
//!   a unit byte (`'s'`/`'u'`/`'n'`) and add a sub-second infinint for `'u'`/`'n'`.
//! - **FSA** (format 9+ only): inode flag bit `0x10` (FSA-full) adds inode
//!   infinints and an FSA block; format 8 has no FSA.
//! - **archive_offset**: points *directly* to the raw file bytes, not to the
//!   data-section header that precedes them in the body stream.
//!   `seek(archive_origin + archive_offset)` then `read(stored_size)`.
//!
//! Full format notes: `docs/implementation-notes.md`.

use std::io::{Cursor, Read, Seek, SeekFrom, Write};

use thiserror::Error;

/// `00 00 00 7b` — DAR magic (SAUV_MAGIC_NUMBER = 123, big-endian u32).
const DAR_MAGIC: [u8; 4] = [0x00, 0x00, 0x00, 0x7b];

/// Upper bound on the compressed catalogue bytes read from the archive tail and
/// on the inflated catalogue, guarding against a decompression bomb (per-file
/// streams need no such constant — they are bounded by the entry's known size).
const MAX_CATALOGUE_COMPRESSED: u64 = 512 * 1024 * 1024;
const MAX_CATALOGUE_INFLATED: u64 = 1024 * 1024 * 1024;

/// Escape sequence marking the catalog: `AD FD EA 77 21 43`.
const SEQT_CATALOGUE: [u8; 6] = [0xAD, 0xFD, 0xEA, 0x77, 0x21, 0x43];

/// First archive format with an in-place (working-directory) path in the
/// catalog header — `archive_version(11,1)` → `value() = 11*256 + 1`.
/// Formats 8, 9, 10 and 11.0 have no such field.
const FORMAT_11_1: u32 = 11 * 256 + 1;

/// Errors returned by [`DarReader`].
#[derive(Debug, Error)]
pub enum DarError {
    #[error("I/O error: {0}")]
    Io(#[from] std::io::Error),
    #[error("not a DAR archive")]
    NotADar,
    #[error("corrupt archive: {0}")]
    Corrupt(String),
    #[error("entry not found: '{0}'")]
    EntryNotFound(String),
}

/// Metadata about one archived file.
#[derive(Debug, Clone)]
pub struct DarEntry {
    pub path: String,
    pub size: u64,
}

#[derive(Debug, Clone)]
struct EntryRef {
    path: String,
    size: u64,
    archive_offset: u64,
    stored_size: u64,
    compression: u8,
    encrypted: bool,
}

/// Read-only DAR archive reader.
pub struct DarReader<R: Read + Seek> {
    inner: R,
    /// Byte position immediately after the slice header TLV block.
    /// `archive_origin + archive_offset` = absolute position of raw file bytes.
    archive_origin: u64,
    entries: Vec<EntryRef>,
}

impl<R: Read + Seek> DarReader<R> {
    /// Open a DAR archive, validating the magic and loading the catalog.
    pub fn open(mut reader: R) -> Result<Self, DarError> {
        let mut magic = [0u8; 4];
        reader
            .read_exact(&mut magic)
            .map_err(|_| DarError::NotADar)?;
        if magic != DAR_MAGIC {
            return Err(DarError::NotADar);
        }

        let mut label = [0u8; 10];
        reader.read_exact(&mut label)?; // internal_name label
        let _flag = read_u8(&mut reader)?; // slice flag ('T' terminal / 'N' / 'E')
        let extension = read_u8(&mut reader)?; // 'T' = TLV (format 8+); 'N'/'S' = legacy (<= 7)

        // Format 8+ carries a TLV list and a `seqt_catalogue` escape; format <= 7
        // has neither — its catalogue is located via the end `terminateur` trailer
        // (libdar header.cpp extension handling; terminateur.cpp).
        let entries;
        let archive_origin;
        if extension == b'T' {
            // TLV list: infinint(count) then count × (u16 type + infinint len + data)
            let tlv_count = read_infinint(&mut reader).map_err(|e| match e {
                DarError::Io(_) => DarError::Corrupt("truncated TLV block".into()),
                other => other,
            })?;
            for _ in 0..tlv_count {
                skip(&mut reader, 2)?;
                let len = read_infinint(&mut reader)?;
                skip(&mut reader, len)?;
            }

            archive_origin = reader.stream_position()?;
            let format_value = read_format_value(&mut reader);
            // The archive's global compression algorithm is the byte immediately
            // after the version string; it tells us whether (and how) the
            // catalogue stream is compressed. Unreadable → treat as stored.
            let global_comp = read_u8(&mut reader).unwrap_or(b'n');
            reader.seek(SeekFrom::Start(archive_origin))?;

            // true → seqt_catalogue tape mark found (catalog has label + maybe path);
            // false → located by its ref_data_name label (tape marks off, e.g. Passware).
            let via_escape = find_catalogue(&mut reader, &label)?;
            let format_major = format_value >> 8;
            if via_escape && is_compressed(global_comp) {
                // The catalogue is a single stream compressed with the archive
                // codec, beginning right after the seqt_catalogue escape and
                // running to the trailer. Inflate it, then parse from the
                // plaintext buffer — which begins with the in-catalog label and
                // optional in-place path, exactly like the uncompressed case.
                let mut compressed = Vec::new();
                reader
                    .by_ref()
                    .take(MAX_CATALOGUE_COMPRESSED)
                    .read_to_end(&mut compressed)?;
                let inflated = decompress(&compressed, global_comp, MAX_CATALOGUE_INFLATED)?;
                let mut cur = Cursor::new(inflated);
                skip(&mut cur, 10)?; // catalog label
                if format_value >= FORMAT_11_1 {
                    skip_nul_string(&mut cur)?;
                }
                entries = parse_catalog(&mut cur, format_major)?;
            } else {
                if via_escape {
                    skip(&mut reader, 10)?; // catalog label
                                            // The in-place path exists only from format 11.1
                                            // (catalogue.cpp:157). Formats 8/9/10/11.0 have none.
                    if format_value >= FORMAT_11_1 {
                        skip_nul_string(&mut reader)?;
                    }
                }
                entries = parse_catalog(&mut reader, format_major)?;
            }
        } else if extension == b'N' || extension == b'S' {
            if extension == b'S' {
                read_infinint(&mut reader)?; // slice size (multi-slice header); unused
            }
            archive_origin = reader.stream_position()?;
            let format_value = read_format_value(&mut reader); // 3-byte edition: value = major*256
            let cat_offset = read_terminateur(&mut reader)?;
            let cat_start = archive_origin
                .checked_add(cat_offset)
                .ok_or_else(|| DarError::Corrupt("catalogue offset overflows".into()))?;
            let end = reader.seek(SeekFrom::End(0))?;
            if cat_start >= end {
                return Err(DarError::Corrupt(format!(
                    "catalogue start {cat_start} past archive end {end}"
                )));
            }
            reader.seek(SeekFrom::Start(cat_start))?;
            // Legacy catalogue: no 10-byte label, no path — entries begin here.
            entries = parse_catalog(&mut reader, format_value >> 8)?;
        } else {
            return Err(DarError::Corrupt(format!(
                "unknown slice-header extension {extension:#04x}"
            )));
        }

        Ok(Self {
            inner: reader,
            archive_origin,
            entries,
        })
    }

    /// List all archived file entries (path and uncompressed size).
    pub fn entries(&self) -> Vec<DarEntry> {
        self.entries
            .iter()
            .map(|e| DarEntry {
                path: e.path.clone(),
                size: e.size,
            })
            .collect()
    }

    /// Extract a file by path, returning its raw bytes.
    pub fn extract(&mut self, path: &str) -> Result<Vec<u8>, DarError> {
        let entry = self
            .entries
            .iter()
            .find(|e| e.path == path)
            .ok_or_else(|| DarError::EntryNotFound(path.to_string()))?
            .clone();

        if entry.encrypted {
            return Err(DarError::Corrupt(format!("'{path}' is encrypted")));
        }

        // The raw bytes live at archive_origin + archive_offset.  Both fields
        // are attacker-controlled, so the sum must be checked, and the claimed
        // size validated against the bytes that actually exist before any
        // allocation — otherwise a forged stored_size is an allocation bomb.
        let start = self
            .archive_origin
            .checked_add(entry.archive_offset)
            .ok_or_else(|| {
                DarError::Corrupt(format!("'{path}' archive offset overflows file position"))
            })?;
        let end = self.inner.seek(SeekFrom::End(0))?;
        if start > end {
            return Err(DarError::Corrupt(format!(
                "'{path}' starts at {start}, past archive end {end}"
            )));
        }
        let available = end - start;
        if entry.stored_size > available {
            return Err(DarError::Corrupt(format!(
                "'{path}' claims {} stored bytes but only {available} remain",
                entry.stored_size
            )));
        }

        self.inner.seek(SeekFrom::Start(start))?;
        let mut data = vec![0u8; entry.stored_size as usize];
        self.inner.read_exact(&mut data)?;

        if !is_compressed(entry.compression) {
            return Ok(data);
        }
        // Each compressed entry is an independent stream; its uncompressed length
        // is the catalog `size`, so decode exactly that and reject any mismatch —
        // a forged stream cannot over-inflate past the declared size.
        let out = decompress(&data, entry.compression, entry.size)?;
        if out.len() as u64 != entry.size {
            return Err(DarError::Corrupt(format!(
                "'{path}' decompressed to {} bytes but catalog declares {}",
                out.len(),
                entry.size
            )));
        }
        Ok(out)
    }
}

// ── Catalog parser ────────────────────────────────────────────────────────────

/// On archives larger than this, the catalog scan starts this many bytes
/// before EOF (the catalog always lives at the tail), avoiding a full read of
/// a multi-gigabyte forensic archive before falling back to a full scan.
const TAIL_SCAN: u64 = 256 * 1024 * 1024;

const CHUNK: usize = 4 * 1024 * 1024;
// OVERLAP = max(SEQT_CATALOGUE.len(), label.len()) - 1; carries bytes across chunk boundaries.
const OVERLAP: usize = 9;

/// Scan forward from the current reader position searching for either the
/// `seqt_catalogue` escape or the archive `label`.
///
/// Returns `Some(true)` if the escape was found (reader positioned just after it),
/// `Some(false)` if the label was found (reader positioned just after it),
/// `None` if EOF was reached without a match.
fn scan_window<R: Read + Seek>(
    r: &mut R,
    label: &[u8; 10],
    use_label: bool,
) -> Result<Option<bool>, DarError> {
    let mut buf = vec![0u8; CHUNK + OVERLAP];
    let mut overlap_len: usize = 0;
    loop {
        let chunk_file_pos = r.stream_position()?;
        let n = r.read(&mut buf[overlap_len..overlap_len + CHUNK])?;
        if n == 0 {
            break;
        }
        let total = overlap_len + n;
        // buf[0..overlap_len]  → tail of previous chunk (file pos: chunk_file_pos - overlap_len)
        // buf[overlap_len..total] → newly read bytes
        let buf_base = chunk_file_pos - overlap_len as u64;

        if let Some(i) = buf[..total]
            .windows(SEQT_CATALOGUE.len())
            .position(|w| w == SEQT_CATALOGUE)
        {
            r.seek(SeekFrom::Start(
                buf_base + i as u64 + SEQT_CATALOGUE.len() as u64,
            ))?;
            return Ok(Some(true));
        }
        if use_label {
            if let Some(i) = buf[..total]
                .windows(label.len())
                .position(|w| w == label.as_ref())
            {
                r.seek(SeekFrom::Start(buf_base + i as u64 + label.len() as u64))?;
                return Ok(Some(false));
            }
        }

        let keep = OVERLAP.min(total);
        buf.copy_within(total - keep..total, 0);
        overlap_len = keep;
    }
    Ok(None)
}

/// Locate the catalog section and position the reader at its first entry.
///
/// Returns `true` when the `seqt_catalogue` escape is found — the caller then
/// skips the 10-byte in-catalog label and (format 11.1+) the path NUL string.
/// The escape is a *sequential-read tape mark*; it is present only when the
/// archive was written with tape marks (libdar's default).
///
/// Returns `false` when the catalog is located by its `ref_data_name` label
/// directly. Archives written with tape marks disabled (e.g. by Passware Kit
/// Mobile, equivalent to `dar -at`) omit the escape; their catalog still begins
/// with the 10-byte `ref_data_name`, which equals the slice `label`, so scanning
/// for `label` in the tail finds it — a structural marker, not a heuristic.
///
/// Returns `Err(Corrupt)` when neither marker is found.
///
/// Strategy: DAR catalogs always live at the tail of the archive.  On forensic
/// archives ≥ 256 MiB we jump straight to the last 256 MiB and scan forward
/// from there, then fall back to a full forward scan from `archive_origin` if
/// needed.  This reduces the I/O for a 92 GiB archive from ~99 GiB to ~107 MiB.
fn find_catalogue<R: Read + Seek>(r: &mut R, label: &[u8; 10]) -> Result<bool, DarError> {
    find_catalogue_within(r, label, TAIL_SCAN)
}

/// Implementation of [`find_catalogue`] with the tail-scan window size as a
/// parameter so the full-scan fallback can be exercised without a 256 MiB
/// fixture.
fn find_catalogue_within<R: Read + Seek>(
    r: &mut R,
    label: &[u8; 10],
    tail_scan: u64,
) -> Result<bool, DarError> {
    // All-zero labels cannot be used as a reliable catalog marker (too common
    // in zero-padded archive bodies).
    let use_label = !label.iter().all(|&b| b == 0);

    let archive_origin = r.stream_position()?;
    let file_end = r.seek(SeekFrom::End(0))?;

    if file_end <= archive_origin {
        return Err(DarError::Corrupt("archive body too short".into()));
    }

    // Jump to at most tail_scan bytes before end; for small files this equals archive_origin.
    let tail_start = archive_origin.max(file_end.saturating_sub(tail_scan));
    r.seek(SeekFrom::Start(tail_start))?;

    if let Some(result) = scan_window(r, label, use_label)? {
        return Ok(result);
    }

    // Tail scan missed.  Fall back to a full scan from archive_origin.
    if tail_start > archive_origin {
        r.seek(SeekFrom::Start(archive_origin))?;
        if let Some(result) = scan_window(r, label, use_label)? {
            return Ok(result);
        }
    }

    Err(DarError::Corrupt("seqt_catalogue not found".into()))
}

/// Read the NUL-terminated `version_string` at the current position and return
/// `archive_version::value()` = `major*256 + fix`, where `major = b0*256 + b1`
/// and each byte is `value + 48`. Format <= 7 stores only `"NN"` (fix implicitly
/// 0); format 8+ stores `"NNf"`. Returns `u32::MAX` for an unreadable string so
/// an unknown future format is treated as newest.
fn read_format_value<R: Read>(r: &mut R) -> u32 {
    let s = read_nul_string(r).unwrap_or_default();
    let b = s.as_bytes();
    if b.len() >= 2 {
        let major = u32::from(b[0].saturating_sub(48)) * 256 + u32::from(b[1].saturating_sub(48));
        let fix = if b.len() >= 3 {
            u32::from(b[2].saturating_sub(48))
        } else {
            0
        };
        major * 256 + fix
    } else {
        u32::MAX
    }
}

/// True when a libdar compression char names a known compression algorithm.
/// `compression2char` emits the algorithm letter in lowercase for streamed mode
/// and uppercase for per-block mode (`z`=gzip, `y`=bzip2, `x`=xz, `l`/`j`/`k`=lzo
/// variants, `d`=zstd, `q`=lz4); `n` is stored. Any other byte — e.g. a header
/// placeholder in a non-dar-produced archive — is treated as not compressed, so
/// the catalogue/entry is read verbatim rather than mis-decoded.
fn is_compressed(algo: u8) -> bool {
    matches!(
        algo.to_ascii_lowercase(),
        b'z' | b'y' | b'x' | b'l' | b'j' | b'k' | b'd' | b'q'
    )
}

/// Inflate one compressed stream, dispatching on the libdar codec char and
/// rejecting output longer than `max_out` (decompression-bomb guard). Trailing
/// bytes after the stream (e.g. the archive trailer) are ignored by the decoder.
fn decompress(data: &[u8], algo: u8, max_out: u64) -> Result<Vec<u8>, DarError> {
    match algo.to_ascii_lowercase() {
        // dar's "gzip" is a raw zlib stream (78 xx), not a gzip (1f 8b) wrapper.
        b'z' => read_bounded(flate2::read::ZlibDecoder::new(data), max_out, "zlib"),
        b'y' => read_bounded(bzip2_rs::DecoderReader::new(data), max_out, "bzip2"),
        b'x' => {
            // lzma-rs is writer-driven and has no output cap, so a BoundedWriter
            // enforces the same decompression-bomb guard the Read codecs get.
            let mut input: &[u8] = data;
            let mut out = BoundedWriter {
                buf: Vec::new(),
                max: max_out,
            };
            match lzma_rs::xz_decompress(&mut input, &mut out) {
                Ok(()) => Ok(out.buf),
                // The DAR trailer follows the catalogue's xz stream. lzma-rs
                // fully decodes and validates the stream (blocks, index, CRC,
                // footer magic) before rejecting trailing bytes, so on this one
                // error the output is already complete and sound. Per-file
                // extract passes exactly stored_size bytes and never trails.
                // (String coupling is why lzma-rs is pinned to 0.3.x.)
                Err(lzma_rs::error::Error::XzError(ref m))
                    if m == "Unexpected data after last XZ block" =>
                {
                    Ok(out.buf)
                }
                Err(e) => Err(DarError::Corrupt(format!("xz decode failed: {e}"))),
            }
        }
        other => Err(DarError::Corrupt(format!(
            "unsupported compression '{}'",
            other as char
        ))),
    }
}

/// A `Write` sink that buffers up to `max` bytes and then fails, capping the
/// output of a writer-driven decoder (lzma-rs) against a decompression bomb.
struct BoundedWriter {
    buf: Vec<u8>,
    max: u64,
}

impl Write for BoundedWriter {
    fn write(&mut self, data: &[u8]) -> std::io::Result<usize> {
        if self.buf.len() as u64 + data.len() as u64 > self.max {
            return Err(std::io::Error::other("decompressed data exceeds bound"));
        }
        self.buf.extend_from_slice(data);
        Ok(data.len())
    }

    fn flush(&mut self) -> std::io::Result<()> {
        Ok(())
    }
}

/// Read a decoder to EOF, capping output at `max_out` bytes (one extra byte is
/// requested so an over-long stream is detected, not silently truncated).
fn read_bounded<R: Read>(decoder: R, max_out: u64, what: &str) -> Result<Vec<u8>, DarError> {
    let mut out = Vec::new();
    decoder
        .take(max_out.saturating_add(1))
        .read_to_end(&mut out)
        .map_err(|e| DarError::Corrupt(format!("{what} decode failed: {e}")))?;
    if out.len() as u64 > max_out {
        return Err(DarError::Corrupt("decompressed data exceeds bound".into()));
    }
    Ok(out)
}

/// Locate the catalogue in a pre-format-8 archive via the end `terminateur`
/// trailer (libdar terminateur.cpp:95-138), returning the catalogue start offset
/// relative to `archive_origin`.
///
/// From EOF, count trailing `0xFF` padding bytes (8 bits each); the first
/// non-`0xFF` byte encodes the remaining count in unary as its set high bits.
/// `byte_offset = total_bits * 4` is the distance back from that byte to the
/// catalogue-position infinint. The `0xFF` run is bounded so a hostile all-`0xFF`
/// tail cannot spin or overflow.
fn read_terminateur<R: Read + Seek>(r: &mut R) -> Result<u64, DarError> {
    const BLOCK_SIZE: u64 = 4;
    const MAX_BITS: u64 = 4096; // far beyond any real terminator

    let mut pos = r.seek(SeekFrom::End(0))?;
    let mut bits: u64 = 0;
    let terminal = loop {
        if pos == 0 {
            return Err(DarError::Corrupt("terminator underflows archive".into()));
        }
        pos -= 1;
        r.seek(SeekFrom::Start(pos))?;
        let b = read_u8(r)?;
        if b == 0xFF {
            bits += 8;
            if bits > MAX_BITS {
                return Err(DarError::Corrupt("terminator padding too long".into()));
            }
        } else {
            break b;
        }
    };
    // The terminator byte must have its top bit set; count consecutive set MSBs.
    if terminal & 0x80 == 0 {
        return Err(DarError::Corrupt(format!(
            "invalid terminator byte {terminal:#04x}"
        )));
    }
    let mut x = terminal;
    while x != 0 {
        if x & 0x80 == 0 {
            return Err(DarError::Corrupt("malformed terminator bit run".into()));
        }
        bits += 1;
        x <<= 1;
    }
    let byte_offset = bits * BLOCK_SIZE;
    let infinint_start = pos
        .checked_sub(byte_offset)
        .ok_or_else(|| DarError::Corrupt("terminator offset underflows".into()))?;
    r.seek(SeekFrom::Start(infinint_start))?;
    read_infinint(r)
}

/// Parse all catalog entries, returning file entries with their extraction info.
///
/// Stops when the root directory is closed (depth reaches zero) or an unknown
/// entry type is encountered (slice trailer).
fn parse_catalog<R: Read + Seek>(r: &mut R, format_major: u32) -> Result<Vec<EntryRef>, DarError> {
    let mut entries = Vec::new();
    let mut dir_stack: Vec<String> = Vec::new();
    let mut depth: u32 = 0;

    loop {
        let mut buf = [0u8; 1];
        match r.read_exact(&mut buf) {
            Ok(()) => {}
            Err(_) => break,
        }

        // Lower 5 bits of cat_sig + 0x60 gives the ASCII type letter.
        let entry_type = ((buf[0] & 0x1f) | 0x60) as char;

        match entry_type {
            'z' => {
                // End of directory
                depth = depth.saturating_sub(1);
                dir_stack.pop();
                if depth == 0 {
                    break;
                }
            }
            'd' => {
                let name = read_nul_string(r)?;
                let flags = read_inode_base(r, format_major)?;
                if format_major >= 9 && (flags >> 4) & 1 != 0 {
                    skip_fsa(r)?;
                }
                depth += 1;
                // <ROOT> is a virtual root; don't include it in file paths.
                if name != "<ROOT>" {
                    dir_stack.push(name);
                }
            }
            'f' => {
                let name = read_nul_string(r)?;
                let flags = read_inode_base(r, format_major)?;
                if format_major >= 9 && (flags >> 4) & 1 != 0 {
                    skip_fsa(r)?;
                }

                let size = read_infinint(r)?;
                let archive_offset = read_infinint(r)?;
                let mut stored_size = read_infinint(r)?;
                // Format <= 7 has no per-file encryption/compression bytes and a
                // fixed 2-byte CRC (no length prefix); format 8+ stores both bytes
                // and a length-prefixed CRC (libdar cat_file.cpp:160-252, crc.cpp).
                let (encryption_flag, compression) = if format_major >= 8 {
                    (read_u8(r)?, read_u8(r)?)
                } else {
                    (0u8, b'n')
                };
                if format_major >= 8 {
                    let crc_size = read_infinint(r)?;
                    skip(r, crc_size)?;
                } else {
                    skip(r, 2)?; // fixed 2-byte CRC
                }
                // Pre-8: storage_size 0 means the data is stored uncompressed.
                if format_major <= 7 && stored_size == 0 {
                    stored_size = size;
                }

                let path = if dir_stack.is_empty() {
                    name
                } else {
                    format!("{}/{}", dir_stack.join("/"), name)
                };

                entries.push(EntryRef {
                    path,
                    size,
                    archive_offset,
                    stored_size,
                    compression,
                    encrypted: encryption_flag != 0,
                });
            }
            'l' => {
                // Symbolic link: inode + NUL-terminated target path; not extractable.
                let _name = read_nul_string(r)?;
                let flags = read_inode_base(r, format_major)?;
                if format_major >= 9 && (flags >> 4) & 1 != 0 {
                    skip_fsa(r)?;
                }
                skip_nul_string(r)?; // symlink target
            }
            _ => break, // unknown type = slice trailer or unhandled entry
        }
    }

    Ok(entries)
}

// ── Low-level I/O helpers ─────────────────────────────────────────────────────

/// Read a DAR variable-length infinint, decoded to `u64`.
///
/// Format (TG=4): optional leading `0x00` skip-bytes, then a terminal byte
/// with exactly one bit set; `pos = terminal.leading_zeros()` and the value
/// occupies `(skip_count * 8 + pos + 1) * 4` big-endian bytes.
///
/// A `u64` holds at most 8 data bytes.  Any encoding wider than that — i.e.
/// *any* leading `0x00` (which alone implies ≥ 36 bytes) or a terminal below
/// `0x40` (`pos > 1`) — cannot be represented and is rejected as `Corrupt`
/// rather than silently truncated.  This single bound also removes the
/// `(skip * 8 …)` arithmetic-overflow panic and caps the leading-zero scan, so
/// a malicious all-zero run can never spin or overflow the skip counter.
fn read_infinint<R: Read>(r: &mut R) -> Result<u64, DarError> {
    let terminal = read_u8(r)?;
    if terminal == 0x00 {
        // A skip-byte group is at least 36 data bytes — far beyond u64.
        return Err(DarError::Corrupt(
            "infinint exceeds 64-bit range (multi-group encoding)".into(),
        ));
    }
    if terminal.count_ones() != 1 {
        return Err(DarError::Corrupt(format!(
            "invalid infinint terminal: {terminal:#04x}"
        )));
    }
    let pos = terminal.leading_zeros(); // 0 ..= 7
    if pos > 1 {
        // data_bytes = (pos + 1) * 4 > 8 → does not fit in u64.
        return Err(DarError::Corrupt(format!(
            "infinint exceeds 64-bit range: terminal {terminal:#04x} implies {} bytes",
            (pos + 1) * 4
        )));
    }
    let data_bytes = (pos + 1) * 4; // 4 (terminal 0x80) or 8 (terminal 0x40)
    let mut val: u64 = 0;
    for _ in 0..data_bytes {
        val = (val << 8) | u64::from(read_u8(r)?);
    }
    Ok(val)
}

fn read_u8<R: Read>(r: &mut R) -> Result<u8, DarError> {
    let mut b = [0u8; 1];
    r.read_exact(&mut b)?;
    Ok(b[0])
}

/// Upper bound on a NUL-terminated path/name field.  Real DAR entries stay
/// well under this; the cap stops a NUL-free region of a hostile archive from
/// growing the buffer until EOF (or OOM on a multi-GiB stream).
const MAX_NUL_STRING: usize = 64 * 1024;

/// Read a NUL-terminated UTF-8 string, consuming the NUL byte.
fn read_nul_string<R: Read>(r: &mut R) -> Result<String, DarError> {
    let mut bytes = Vec::new();
    loop {
        let b = read_u8(r)?;
        if b == 0 {
            break;
        }
        if bytes.len() >= MAX_NUL_STRING {
            return Err(DarError::Corrupt(format!(
                "NUL-terminated string exceeds {MAX_NUL_STRING} bytes"
            )));
        }
        bytes.push(b);
    }
    String::from_utf8(bytes).map_err(|e| DarError::Corrupt(e.to_string()))
}

/// Skip a NUL-terminated string without collecting the bytes.
fn skip_nul_string<R: Read>(r: &mut R) -> Result<(), DarError> {
    let mut len: usize = 0;
    loop {
        if read_u8(r)? == 0 {
            return Ok(());
        }
        len += 1;
        if len > MAX_NUL_STRING {
            return Err(DarError::Corrupt(format!(
                "NUL-terminated string exceeds {MAX_NUL_STRING} bytes"
            )));
        }
    }
}

/// Seek past `n` bytes.
fn skip<R: Seek>(r: &mut R, n: u64) -> Result<(), DarError> {
    if n > 0 {
        // `SeekFrom::Current` takes an i64; a value above i64::MAX would cast to
        // a negative offset and seek *backwards* (re-reading earlier bytes on a
        // File).  No real DAR field is that large — reject it outright.
        let off = i64::try_from(n)
            .map_err(|_| DarError::Corrupt(format!("skip length {n} exceeds seekable range")))?;
        r.seek(SeekFrom::Current(off)).map_err(DarError::Io)?;
    }
    Ok(())
}

/// Skip one DAR timestamp field.
///
/// Timestamps are prefixed with a type byte:
/// - `'s'` (0x73) and others: seconds only — one infinint follows
/// - `'n'` (0x6e): nanosecond precision — two infinints follow (seconds + nanoseconds)
fn skip_timestamp<R: Read + Seek>(r: &mut R, format_major: u32) -> Result<(), DarError> {
    // Format 8 and earlier store a bare seconds infinint with NO precision byte
    // (libdar datetime.cpp:372). Format 9+ prefix a unit byte ('s' seconds,
    // 'u' microsecond, 'n' nanosecond); sub-second units add a second infinint.
    if format_major < 9 {
        read_infinint(r)?;
        return Ok(());
    }
    let ts_type = read_u8(r)?;
    read_infinint(r)?;
    if ts_type == b'n' || ts_type == b'u' {
        read_infinint(r)?;
    }
    Ok(())
}

/// Read the inode flags byte and seek past the remaining inode fields.
///
/// Base layout: flags(1) + uid(inf) + gid(inf) + perms(2) + 3 timestamps
///   Each timestamp: see [`skip_timestamp`] (version-dependent).
///   FSA inode fields (format 9+ only): two infinints when (flags >> 4) & 1 == 1.
fn read_inode_base<R: Read + Seek>(r: &mut R, format_major: u32) -> Result<u8, DarError> {
    let flags = read_u8(r)?;
    // uid/gid: 2-byte u16 for format <= 7 (libdar cat_inode.cpp:171), infinint for 8+.
    if format_major <= 7 {
        skip(r, 4)?; // uid (u16) + gid (u16)
    } else {
        read_infinint(r)?; // uid
        read_infinint(r)?; // gid
    }
    skip(r, 2)?; // perms (always a 2-byte big-endian u16, never an infinint)
    skip_timestamp(r, format_major)?; // atime
    skip_timestamp(r, format_major)?; // mtime
                                      // ctime (last_cha) exists only from format 8 (libdar cat_inode.cpp:197).
    if format_major >= 8 {
        skip_timestamp(r, format_major)?;
    }
    // FSA inode fields exist only from format 9 (libdar cat_inode.cpp:264); bit
    // 0x10 is the FSA-full status. Formats <= 8 have no FSA.
    if format_major >= 9 && (flags >> 4) & 1 != 0 {
        read_infinint(r)?;
        read_infinint(r)?;
    }
    Ok(flags)
}

/// Skip one FSA (filesystem attributes) block.
///
/// Format: infinint(family_tag) + infinint(data_size) + data_size bytes.
fn skip_fsa<R: Read + Seek>(r: &mut R) -> Result<(), DarError> {
    let _tag = read_infinint(r)?;
    let size = read_infinint(r)?;
    skip(r, size)
}

// ── Unit tests ────────────────────────────────────────────────────────────────

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

    // ── read_infinint ─────────────────────────────────────────────────────────

    #[test]
    fn infinint_decodes_value() {
        let data = [0x80u8, 0x00, 0x00, 0x00, 0x0d];
        assert_eq!(read_infinint(&mut Cursor::new(&data[..])).unwrap(), 13);
    }

    #[test]
    fn infinint_bad_preamble_returns_corrupt() {
        // 0x03 = two bits set — not a valid infinint terminal.
        let data = [0x03u8, 0x00, 0x00, 0x00, 0x00];
        let err = read_infinint(&mut Cursor::new(&data[..])).unwrap_err();
        assert!(matches!(&err, DarError::Corrupt(_)));
    }

    #[test]
    fn infinint_truncated_returns_io() {
        // Only 2 bytes — read_exact needs 5.
        let err = read_infinint(&mut Cursor::new(&[0x80u8, 0x00][..])).unwrap_err();
        assert!(matches!(err, DarError::Io(_)));
    }

    #[test]
    fn infinint_0x40_preamble_reads_8_data_bytes() {
        // 0x40 terminal: leading_zeros=1, pos=1, data_bytes=(0*8+1+1)*4=8
        // Encodes the value 0x5d15_9331 in 8 big-endian bytes.
        let mut data = vec![0x40u8];
        data.extend_from_slice(&[0x00, 0x00, 0x00, 0x00, 0x5d, 0x15, 0x93, 0x31]);
        assert_eq!(
            read_infinint(&mut Cursor::new(data)).unwrap(),
            0x5d15_9331u64
        );
    }

    #[test]
    fn infinint_multi_bit_terminal_returns_corrupt() {
        // 0x60 = 0110_0000 — two bits set, not a valid terminal.
        let data = [0x60u8, 0x00, 0x00, 0x00, 0x00];
        let err = read_infinint(&mut Cursor::new(&data[..])).unwrap_err();
        assert!(matches!(&err, DarError::Corrupt(_)));
    }

    // ── read_u8 ───────────────────────────────────────────────────────────────

    #[test]
    fn read_u8_reads_single_byte() {
        assert_eq!(read_u8(&mut Cursor::new(&[0x42u8][..])).unwrap(), 0x42);
    }

    #[test]
    fn read_u8_eof_returns_io() {
        let err = read_u8(&mut Cursor::new(&[][..])).unwrap_err();
        assert!(matches!(err, DarError::Io(_)));
    }

    // ── read_nul_string ───────────────────────────────────────────────────────

    #[test]
    fn nul_string_reads_until_nul() {
        let data = b"hello\x00world";
        assert_eq!(
            read_nul_string(&mut Cursor::new(&data[..])).unwrap(),
            "hello"
        );
    }

    #[test]
    fn nul_string_invalid_utf8_returns_corrupt() {
        // 0xFF 0x80 is not valid UTF-8; 0x00 terminates.
        let data = [0xFF, 0x80, 0x00];
        let err = read_nul_string(&mut Cursor::new(&data[..])).unwrap_err();
        assert!(matches!(err, DarError::Corrupt(_)));
    }

    #[test]
    fn nul_string_eof_before_nul_returns_io() {
        let err = read_nul_string(&mut Cursor::new(b"no-nul".to_vec())).unwrap_err();
        assert!(matches!(err, DarError::Io(_)));
    }

    // ── skip_nul_string ───────────────────────────────────────────────────────

    #[test]
    fn skip_nul_string_advances_past_nul() {
        let data = b"skip\x00rest";
        let mut c = Cursor::new(data.to_vec());
        skip_nul_string(&mut c).unwrap();
        assert_eq!(c.position(), 5); // "skip\0" = 5 bytes consumed
    }

    #[test]
    fn skip_nul_string_eof_returns_io() {
        let err = skip_nul_string(&mut Cursor::new(b"no-nul".to_vec())).unwrap_err();
        assert!(matches!(err, DarError::Io(_)));
    }

    // ── find_catalogue ────────────────────────────────────────────────────────

    #[test]
    fn find_catalogue_body_too_short() {
        // Fewer than 6 bytes — can't fill the initial window; label also too short.
        let label = [0u8; 10];
        let err = find_catalogue(&mut Cursor::new(&[0x01u8, 0x02, 0x03][..]), &label).unwrap_err();
        assert!(
            matches!(&err, DarError::Corrupt(s) if s == "archive body too short"
            || s == "seqt_catalogue not found")
        );
    }

    #[test]
    fn find_catalogue_escape_at_start() {
        let mut data = [0xAD, 0xFD, 0xEA, 0x77, 0x21, 0x43, 0xFF];
        let mut c = Cursor::new(&mut data[..]);
        let via_escape = find_catalogue(&mut c, &[0u8; 10]).unwrap();
        assert!(via_escape);
        assert_eq!(c.position(), 6);
    }

    #[test]
    fn find_catalogue_escape_not_found() {
        // 10 bytes of zeros, label is 0xFF×10 so label scan also fails.
        let label = [0xFFu8; 10];
        let err = find_catalogue(&mut Cursor::new(&[0u8; 10][..]), &label).unwrap_err();
        assert!(matches!(&err, DarError::Corrupt(s) if s == "seqt_catalogue not found"));
    }

    #[test]
    fn find_catalogue_label_fallback() {
        let label: [u8; 10] = [0xA1, 0xB2, 0xC3, 0xD4, 0xE5, 0xF6, 0x07, 0x18, 0x29, 0x3A];
        // Prefix junk (no escape) followed by the label bytes.
        let mut data = vec![0x00u8; 5];
        data.extend_from_slice(&label);
        let mut c = Cursor::new(data);
        let via_escape = find_catalogue(&mut c, &label).unwrap();
        assert!(!via_escape);
        assert_eq!(c.position(), 15); // 5 junk + 10 label consumed
    }

    // ── skip ──────────────────────────────────────────────────────────────────

    #[test]
    fn skip_zero_does_not_move_cursor() {
        let mut c = Cursor::new(vec![0xFFu8; 10]);
        skip(&mut c, 0).unwrap();
        assert_eq!(c.position(), 0);
    }

    #[test]
    fn skip_n_advances_cursor() {
        let mut c = Cursor::new(vec![0xFFu8; 10]);
        skip(&mut c, 7).unwrap();
        assert_eq!(c.position(), 7);
    }

    // ── read_inode_base ───────────────────────────────────────────────────────

    #[test]
    fn inode_base_bit4_clear_reads_31_bytes() {
        // flags(1) + uid(5) + gid(5) + perms(2) + 3×[type(1)+secs(5)] = 31 bytes
        let mut data = vec![0x00u8]; // flags (bit4=0)
        data.extend_from_slice(&[0x80, 0x00, 0x00, 0x00, 0x00]); // uid
        data.extend_from_slice(&[0x80, 0x00, 0x00, 0x00, 0x00]); // gid
        data.extend_from_slice(&[0x00, 0x00]); // perms
        for _ in 0..3 {
            data.push(b's'); // timestamp type
            data.extend_from_slice(&[0x80, 0x00, 0x00, 0x00, 0x00]); // seconds
        }
        data.push(0xFF); // sentinel — must not be consumed
        let mut c = Cursor::new(data);
        assert_eq!(read_inode_base(&mut c, 11).unwrap(), 0x00);
        assert_eq!(c.position(), 31);
    }

    #[test]
    fn inode_base_bit4_set_reads_41_bytes() {
        // flags(1) + uid(5) + gid(5) + perms(2) + 3×[type(1)+secs(5)] + nlink(5) + field9(5) = 41
        let mut data = vec![0x10u8]; // flags (bit4=1)
        data.extend_from_slice(&[0x80, 0x00, 0x00, 0x00, 0x00]); // uid
        data.extend_from_slice(&[0x80, 0x00, 0x00, 0x00, 0x00]); // gid
        data.extend_from_slice(&[0x00, 0x00]); // perms
        for _ in 0..3 {
            data.push(b's');
            data.extend_from_slice(&[0x80, 0x00, 0x00, 0x00, 0x00]);
        }
        data.extend_from_slice(&[0x80, 0x00, 0x00, 0x00, 0x00]); // nlink
        data.extend_from_slice(&[0x80, 0x00, 0x00, 0x00, 0x00]); // field9
        data.push(0xFF); // sentinel
        let mut c = Cursor::new(data);
        assert_eq!(read_inode_base(&mut c, 11).unwrap(), 0x10);
        assert_eq!(c.position(), 41);
    }

    // ── skip_fsa ─────────────────────────────────────────────────────────────

    #[test]
    fn skip_fsa_consumes_tag_size_and_data() {
        // tag=infinint(5) + size=infinint(3) + 3 data bytes
        let mut data = Vec::new();
        data.extend_from_slice(&[0x80, 0x00, 0x00, 0x00, 0x05]); // tag
        data.extend_from_slice(&[0x80, 0x00, 0x00, 0x00, 0x03]); // size=3
        data.extend_from_slice(&[0xAA, 0xBB, 0xCC]); // data
        data.push(0xFF); // sentinel
        let mut c = Cursor::new(data);
        skip_fsa(&mut c).unwrap();
        assert_eq!(c.position(), 13); // 5 + 5 + 3 = 13
    }

    // ── hardening: malicious / corrupted infinint encodings ───────────────────
    //
    // A `u64` holds at most 8 data bytes.  The reader's contract is "decode to
    // u64 or return Corrupt" — it must never silently truncate an over-wide
    // value, overflow while computing the byte count, or loop on a zero run.

    #[test]
    fn infinint_leading_zero_byte_returns_corrupt() {
        // A leading 0x00 skip-byte implies a ≥36-byte group — far beyond u64.
        // Must be rejected as Corrupt, not mislabelled as an I/O shortage.
        let data = [0x00u8, 0x80, 0x00, 0x00, 0x00, 0x00];
        let err = read_infinint(&mut Cursor::new(&data[..])).unwrap_err();
        assert!(matches!(err, DarError::Corrupt(_)), "got {err:?}");
    }

    #[test]
    fn infinint_12_byte_group_exceeds_u64_returns_corrupt() {
        // 0x20 terminal → pos=2 → 12 data bytes → cannot fit in u64.
        // Must error rather than silently truncate to a wrong value.
        let mut data = vec![0x20u8];
        data.extend_from_slice(&[0x11; 12]);
        let err = read_infinint(&mut Cursor::new(data)).unwrap_err();
        assert!(matches!(err, DarError::Corrupt(_)), "got {err:?}");
    }

    #[test]
    fn infinint_all_zero_run_returns_corrupt_without_hanging() {
        // A run of zero bytes must terminate promptly with Corrupt, never spin
        // consuming the whole stream (and never overflow-panic the skip count).
        let data = vec![0u8; 4096];
        let err = read_infinint(&mut Cursor::new(data)).unwrap_err();
        assert!(matches!(err, DarError::Corrupt(_)), "got {err:?}");
    }

    // ── hardening: unbounded NUL-terminated strings ───────────────────────────

    #[test]
    fn nul_string_without_terminator_is_length_bounded() {
        // No NUL in 200 KiB of data: must be rejected once the path cap is hit,
        // not grow the buffer until EOF (or OOM on a multi-GiB stream).
        let data = vec![b'A'; 200_000];
        let err = read_nul_string(&mut Cursor::new(data)).unwrap_err();
        assert!(matches!(err, DarError::Corrupt(_)), "got {err:?}");
    }

    #[test]
    fn skip_nul_string_without_terminator_is_length_bounded() {
        let data = vec![b'A'; 200_000];
        let err = skip_nul_string(&mut Cursor::new(data)).unwrap_err();
        assert!(matches!(err, DarError::Corrupt(_)), "got {err:?}");
    }

    // ── hardening: skip must never seek backwards ─────────────────────────────

    #[test]
    fn skip_value_above_i64_max_returns_corrupt() {
        // n > i64::MAX casts to a negative i64 → SeekFrom::Current would seek
        // *backwards* on a File (re-reading earlier bytes).  Must be rejected,
        // and the stream position must not move.
        let mut c = Cursor::new(vec![0u8; 64]);
        c.set_position(32);
        let err = skip(&mut c, 0x8000_0000_0000_0000).unwrap_err();
        assert!(matches!(err, DarError::Corrupt(_)), "got {err:?}");
        assert_eq!(c.position(), 32); // unchanged on a rejected skip
    }

    // ── terminateur trailer (pre-8 catalog locator) ───────────────────────────

    #[test]
    fn terminateur_reads_catalogue_offset() {
        // pos infinint 0x18 = 24; terminator 0xc0 → two leading ones → 2*4 = 8
        // bytes back to the infinint.
        let data = vec![0x80u8, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0xc0];
        assert_eq!(read_terminateur(&mut Cursor::new(data)).unwrap(), 24);
    }

    #[test]
    fn terminateur_all_ff_underflows_returns_corrupt() {
        let err = read_terminateur(&mut Cursor::new(vec![0xFFu8; 4])).unwrap_err();
        assert!(matches!(err, DarError::Corrupt(_)), "got {err:?}");
    }

    #[test]
    fn terminateur_excessive_ff_padding_returns_corrupt() {
        let err = read_terminateur(&mut Cursor::new(vec![0xFFu8; 600])).unwrap_err();
        assert!(matches!(err, DarError::Corrupt(_)), "got {err:?}");
    }

    #[test]
    fn terminateur_low_terminator_byte_returns_corrupt() {
        // Terminator byte 0x01 has no top bit set.
        let data = vec![0x80u8, 0x00, 0x00, 0x00, 0x18, 0x01];
        let err = read_terminateur(&mut Cursor::new(data)).unwrap_err();
        assert!(matches!(err, DarError::Corrupt(_)), "got {err:?}");
    }

    #[test]
    fn terminateur_noncontiguous_high_bits_returns_corrupt() {
        // 0xA0 = 1010_0000: top bit set but the high-bit run is not contiguous.
        let data = vec![0x80u8, 0x00, 0x00, 0x00, 0x18, 0xA0];
        let err = read_terminateur(&mut Cursor::new(data)).unwrap_err();
        assert!(matches!(err, DarError::Corrupt(_)), "got {err:?}");
    }

    // ── find_catalogue: full-scan fallback + body-too-short ────────────────────

    #[test]
    fn find_catalogue_falls_back_to_full_scan() {
        // Escape near the start; a tiny tail window misses it, forcing the
        // archive_origin full-scan fallback.
        let mut data = vec![0x11u8, 0x22]; // junk before the escape
        data.extend_from_slice(&SEQT_CATALOGUE);
        data.extend_from_slice(&[0x33u8; 12]); // trailing bytes beyond the tail window
        let mut c = Cursor::new(data);
        let via_escape = find_catalogue_within(&mut c, &[0u8; 10], 4).unwrap();
        assert!(via_escape);
        assert_eq!(c.position(), 2 + SEQT_CATALOGUE.len() as u64);
    }

    #[test]
    fn find_catalogue_full_scan_miss_returns_not_found() {
        // No escape and no matching label anywhere; a tiny tail window forces
        // the full-scan fallback, which also misses → "not found".
        let mut c = Cursor::new(vec![0x11u8; 16]);
        let err = find_catalogue_within(&mut c, &[0xABu8; 10], 4).unwrap_err();
        assert!(matches!(&err, DarError::Corrupt(s) if s == "seqt_catalogue not found"));
    }

    #[test]
    fn find_catalogue_body_too_short_when_origin_at_eof() {
        let mut c = Cursor::new(vec![0u8; 6]);
        c.seek(SeekFrom::Start(6)).unwrap();
        let err = find_catalogue(&mut c, &[0u8; 10]).unwrap_err();
        assert!(matches!(&err, DarError::Corrupt(s) if s == "archive body too short"));
    }

    // ── decompress ─────────────────────────────────────────────────────────────

    #[test]
    fn decompress_rejects_decompression_bomb() {
        use flate2::{write::ZlibEncoder, Compression};
        use std::io::Write;
        let mut enc = ZlibEncoder::new(Vec::new(), Compression::default());
        enc.write_all(&[0u8; 4096]).unwrap();
        let blob = enc.finish().unwrap();
        // Inflates to 4096 bytes but the caller caps output at 16.
        let err = decompress(&blob, b'z', 16).unwrap_err();
        assert!(matches!(&err, DarError::Corrupt(s) if s.contains("exceeds bound")));
    }

    #[test]
    fn decompress_rejects_malformed_zlib() {
        let err = decompress(b"not a zlib stream at all", b'z', 1024).unwrap_err();
        assert!(matches!(&err, DarError::Corrupt(s) if s.contains("zlib decode failed")));
    }

    #[test]
    fn decompress_rejects_malformed_xz() {
        let err = decompress(b"this is not an xz stream", b'x', 1024).unwrap_err();
        assert!(matches!(&err, DarError::Corrupt(s) if s.contains("xz decode failed")));
    }

    #[test]
    fn bounded_writer_caps_output_and_flushes() {
        let mut w = BoundedWriter {
            buf: Vec::new(),
            max: 4,
        };
        assert_eq!(w.write(b"ab").unwrap(), 2); // within bound
        w.flush().unwrap();
        let err = w.write(b"cde").unwrap_err(); // 2 + 3 > 4
        assert_eq!(err.to_string(), "decompressed data exceeds bound");
        assert_eq!(w.buf, b"ab");
    }
}