diff --git a/src/paths_serialization.rs b/src/paths_serialization.rs
index 7f03034f..eef176d4 100644
--- a/src/paths_serialization.rs
+++ b/src/paths_serialization.rs
@@ -23,6 +23,10 @@ mod paths_serialization_nightly;
 #[cfg(feature="nightly")]
 pub use paths_serialization_nightly::*;
 
+#[path="paths_serialization_sort.rs"]
+mod paths_serialization_sort;
+pub use paths_serialization_sort::sort_paths;
+
 /// Statistics from a `serialize` operation
 #[derive(Debug, Clone, Copy)]
 pub struct SerializationStats {
diff --git a/src/paths_serialization_sort.rs b/src/paths_serialization_sort.rs
new file mode 100644
index 00000000..326f9e47
--- /dev/null
+++ b/src/paths_serialization_sort.rs
@@ -0,0 +1,235 @@
+//! External bytewise sorting for unordered `.paths` streams.
+use super::{SerializationStats, for_each_deserialized_path, serialize_paths_from_funcs};
+use std::{
+    fs::{self, File},
+    io::{self, BufReader, BufWriter, Read, Write},
+    path::{Path, PathBuf},
+};
+
+const BUFFER: usize = 16 * 1024;
+
+fn new_run(temp_dir: &Path, next_run: &mut u64) -> io::Result<(PathBuf, File)> {
+    let path = temp_dir.join(format!("pathmap-sort-{next_run}"));
+    let file = File::create_new(&path)?;
+    *next_run += 1;
+    Ok((path, file))
+}
+
+// Runs are uncompressed length-prefixed paths, avoiding repeated compression
+// during merges. These functions share no codec state or in-memory path index.
+fn read_path(source: &mut impl Read, path: &mut Vec<u8>, limit: usize) -> io::Result<bool> {
+    let mut length = [0; 4];
+    loop {
+        match source.read(&mut length[..1]) {
+            Ok(0) => return Ok(false),
+            Ok(_) => break,
+            Err(e) if e.kind() == io::ErrorKind::Interrupted => continue,
+            Err(e) => return Err(e),
+        }
+    }
+    source.read_exact(&mut length[1..])?;
+    let length = u32::from_le_bytes(length) as usize;
+    if length > limit {
+        return Err(io::Error::new(
+            io::ErrorKind::InvalidData,
+            "path exceeds sort memory budget",
+        ));
+    }
+    path.resize(length, 0);
+    source.read_exact(path)?;
+    Ok(true)
+}
+fn write_path(target: &mut impl Write, path: &[u8]) -> io::Result<()> {
+    let length = u32::try_from(path.len())
+        .map_err(|_| io::Error::other("path exceeds .paths length limit"))?;
+    target.write_all(&length.to_le_bytes())?;
+    target.write_all(path)
+}
+
+struct Chunk {
+    bytes: Vec<u8>,
+    entries: Vec<(usize, usize)>,
+    byte_limit: usize,
+    entry_limit: usize,
+}
+impl Chunk {
+    fn new(memory: usize) -> Self {
+        let byte_limit = memory / 3;
+        let entry_limit = memory / 6 / std::mem::size_of::<(usize, usize)>();
+        Self {
+            bytes: Vec::with_capacity(byte_limit),
+            entries: Vec::with_capacity(entry_limit),
+            byte_limit,
+            entry_limit,
+        }
+    }
+    fn fits(&self, path: &[u8]) -> bool {
+        self.bytes.len() + path.len() <= self.byte_limit && self.entries.len() < self.entry_limit
+    }
+    fn push(&mut self, path: &[u8]) {
+        self.entries.push((self.bytes.len(), path.len()));
+        self.bytes.extend_from_slice(path);
+    }
+    fn spill(&mut self, temp_dir: &Path, next_run: &mut u64) -> io::Result<PathBuf> {
+        let bytes = &self.bytes;
+        self.entries
+            .sort_unstable_by(|&(a, al), &(b, bl)| bytes[a..a + al].cmp(&bytes[b..b + bl]));
+        let (path, file) = new_run(temp_dir, next_run)?;
+        let mut writer = BufWriter::with_capacity(BUFFER, file);
+        let mut previous = None;
+        for &(start, len) in &self.entries {
+            let path = &bytes[start..start + len];
+            if previous != Some(path) {
+                write_path(&mut writer, path)?;
+            }
+            previous = Some(path);
+        }
+        writer.flush()?;
+        drop(writer);
+        self.entries.clear();
+        self.bytes.clear();
+        Ok(path)
+    }
+}
+
+fn merge(
+    left: PathBuf,
+    right: PathBuf,
+    temp_dir: &Path,
+    max_path: usize,
+    next_run: &mut u64,
+) -> io::Result<PathBuf> {
+    let mut a = BufReader::with_capacity(BUFFER, File::open(&left)?);
+    let mut b = BufReader::with_capacity(BUFFER, File::open(&right)?);
+    let mut ap = Vec::new();
+    let mut bp = Vec::new();
+    let (path, output) = new_run(temp_dir, next_run)?;
+    let mut writer = BufWriter::with_capacity(BUFFER, output);
+    let mut has_a = read_path(&mut a, &mut ap, max_path)?;
+    let mut has_b = read_path(&mut b, &mut bp, max_path)?;
+    while has_a || has_b {
+        let order = match (has_a, has_b) {
+            (true, true) => ap.cmp(&bp),
+            (true, false) => std::cmp::Ordering::Less,
+            _ => std::cmp::Ordering::Greater,
+        };
+        write_path(&mut writer, if order.is_le() { &ap } else { &bp })?;
+        if order.is_le() {
+            has_a = read_path(&mut a, &mut ap, max_path)?;
+        }
+        if order.is_ge() {
+            has_b = read_path(&mut b, &mut bp, max_path)?;
+        }
+    }
+    writer.flush()?;
+    drop(writer);
+    drop((a, b));
+    fs::remove_file(left)?;
+    fs::remove_file(right)?;
+    Ok(path)
+}
+
+// Binary carry merging: at most one run per level and no unbounded run list.
+fn add_run(
+    mut run: PathBuf,
+    levels: &mut [Option<PathBuf>; 64],
+    temp_dir: &Path,
+    max_path: usize,
+    next_run: &mut u64,
+) -> io::Result<()> {
+    for slot in levels {
+        match slot.take() {
+            None => {
+                *slot = Some(run);
+                return Ok(());
+            }
+            Some(other) => run = merge(other, run, temp_dir, max_path, next_run)?,
+        }
+    }
+    Err(io::Error::other("too many sort runs"))
+}
+
+/// Sort and deduplicate an unordered `.paths` stream without building a PathMap.
+///
+/// `memory_bytes` (at least 1 MiB) budgets the sorting arena, indices and merge
+/// buffers; codec/runtime overhead and the decoder's current path are additional.
+/// Individual paths must fit in 1/16 of the budget (checked after decoding).
+/// The caller supplies an existing scratch directory in `temp_dir`, exclusive to
+/// this sort. Run files are created with `create_new` and removed on success or
+/// error; the caller owns creation and deletion of the directory. Allow up to
+/// twice the uncompressed input size for runs.
+/// The output uses the usual `.paths` format, in strictly increasing byte order.
+pub fn sort_paths<R: Read, W: Write>(
+    source: R,
+    target: &mut W,
+    memory_bytes: usize,
+    temp_dir: &Path,
+) -> io::Result<SerializationStats> {
+    if memory_bytes < 1024 * 1024 {
+        return Err(io::Error::new(
+            io::ErrorKind::InvalidInput,
+            "sort memory budget must be at least 1 MiB",
+        ));
+    }
+    let mut next_run = 0;
+    let result = (|| {
+        let max_path = memory_bytes / 16;
+        let mut chunk = Chunk::new(memory_bytes);
+        let mut levels = std::array::from_fn(|_| None);
+        for_each_deserialized_path(source, |_, path| {
+            if path.len() > max_path {
+                return Err(io::Error::new(
+                    io::ErrorKind::InvalidData,
+                    "path exceeds sort memory budget",
+                ));
+            }
+            if !chunk.fits(path) {
+                add_run(
+                    chunk.spill(temp_dir, &mut next_run)?,
+                    &mut levels,
+                    temp_dir,
+                    max_path,
+                    &mut next_run,
+                )?;
+            }
+            chunk.push(path);
+            Ok(())
+        })?;
+        if !chunk.entries.is_empty() {
+            add_run(
+                chunk.spill(temp_dir, &mut next_run)?,
+                &mut levels,
+                temp_dir,
+                max_path,
+                &mut next_run,
+            )?;
+        }
+        drop(chunk);
+        let mut final_run = None;
+        for run in levels.into_iter().flatten() {
+            final_run = Some(match final_run {
+                None => run,
+                Some(other) => merge(other, run, temp_dir, max_path, &mut next_run)?,
+            });
+        }
+        let mut source: Box<dyn Read> = match &final_run {
+            Some(path) => Box::new(BufReader::with_capacity(BUFFER, File::open(path)?)),
+            None => Box::new(io::empty()),
+        };
+        serialize_paths_from_funcs(
+            target,
+            &mut Vec::new(),
+            |path| read_path(&mut source, path, max_path),
+            |path| Some(path),
+        )
+    })();
+    let mut cleanup = Ok(());
+    for index in 0..next_run {
+        if let Err(error) = fs::remove_file(temp_dir.join(format!("pathmap-sort-{index}"))) {
+            if error.kind() != io::ErrorKind::NotFound {
+                cleanup = Err(error);
+            }
+        }
+    }
+    result.and_then(|stats| cleanup.map(|()| stats))
+}
diff --git a/tests/paths_sort.rs b/tests/paths_sort.rs
new file mode 100644
index 00000000..045b8080
--- /dev/null
+++ b/tests/paths_sort.rs
@@ -0,0 +1,102 @@
+#![cfg(feature = "serialization")]
+use pathmap::paths_serialization::{
+    for_each_deserialized_path, serialize_paths_from_funcs, sort_paths,
+};
+use std::{
+    fs::{self, File},
+    io,
+    path::Path,
+};
+fn write_paths(path: &Path, paths: &[Vec<u8>]) {
+    let mut source = (0usize, paths);
+    serialize_paths_from_funcs(
+        &mut File::create(path).unwrap(),
+        &mut source,
+        |s| {
+            s.0 += 1;
+            Ok(s.0 <= s.1.len())
+        },
+        |s| Some(s.1[s.0 - 1].as_slice()),
+    )
+    .unwrap();
+}
+fn read_paths(path: &Path) -> Vec<Vec<u8>> {
+    let mut paths = vec![];
+    for_each_deserialized_path(File::open(path).unwrap(), |_, p| {
+        paths.push(p.to_vec());
+        Ok(())
+    })
+    .unwrap();
+    paths
+}
+fn sort_file(input: &Path, output: &Path, memory: usize, temp: &Path) -> io::Result<usize> {
+    let mut target = tempfile::NamedTempFile::new_in(temp)?;
+    let count = sort_paths(File::open(input)?, &mut target, memory, temp)?.path_count;
+    target.persist(output).map_err(|e| e.error)?;
+    Ok(count)
+}
+#[test]
+fn external_sort_many_runs_matches_bytewise_set() {
+    let scratch = tempfile::tempdir().unwrap();
+    let input = scratch.path().join("input.upaths");
+    let output = scratch.path().join("output.paths");
+    let mut paths = vec![vec![], vec![0], vec![0, 0], vec![255], vec![]];
+    for n in (0u32..35000).rev() {
+        let mut path = n.to_be_bytes().to_vec();
+        path.extend_from_slice(&[0, 255, 0]);
+        path.extend(std::iter::repeat_n((n % 251) as u8, (n % 123) as usize));
+        paths.push(path.clone());
+        if n % 3 == 0 {
+            paths.push(path);
+        }
+    }
+    write_paths(&input, &paths);
+    paths.sort();
+    paths.dedup();
+    assert_eq!(
+        sort_file(&input, &output, 1 << 20, scratch.path()).unwrap(),
+        paths.len()
+    );
+    assert_eq!(read_paths(&output), paths);
+    // A sort-budget error after multiple spills must clean every temporary run.
+    let before = fs::read(&output).unwrap();
+    paths.push(vec![0; (1 << 16) + 1]);
+    write_paths(&input, &paths);
+    assert!(sort_file(&input, &output, 1 << 20, scratch.path()).is_err());
+    assert_eq!(fs::read(&output).unwrap(), before);
+    assert_eq!(fs::read_dir(scratch.path()).unwrap().count(), 2);
+}
+
+#[test]
+fn sort_empty_duplicates_and_memory_limit() {
+    let scratch = tempfile::tempdir().unwrap();
+    let input = scratch.path().join("input.upaths");
+    let output = scratch.path().join("output.paths");
+    for paths in [vec![], vec![vec![]; 3], vec![b"same".to_vec(); 25000]] {
+        write_paths(&input, &paths);
+        let mut expected = paths;
+        expected.sort();
+        expected.dedup();
+        sort_file(&input, &output, 1 << 20, scratch.path()).unwrap();
+        assert_eq!(read_paths(&output), expected);
+    }
+    write_paths(&input, &[vec![0; (1 << 16) + 1]]);
+    assert!(sort_file(&input, &output, 1 << 20, scratch.path()).is_err());
+    assert_eq!(fs::read_dir(scratch.path()).unwrap().count(), 2);
+}
+
+#[test]
+fn caller_scratch_files_survive_run_name_collisions() {
+    let scratch = tempfile::tempdir().unwrap();
+    let input = scratch.path().join("input.upaths");
+    let output = scratch.path().join("output.paths");
+    let existing = scratch.path().join("pathmap-sort-1");
+    fs::write(&existing, b"caller-owned file").unwrap();
+    fs::write(&output, b"existing destination").unwrap();
+    write_paths(&input, &vec![vec![42; 64]; 20000]);
+    let error = sort_file(&input, &output, 1 << 20, scratch.path()).unwrap_err();
+    assert_eq!(error.kind(), io::ErrorKind::AlreadyExists);
+    assert_eq!(fs::read(&existing).unwrap(), b"caller-owned file");
+    assert_eq!(fs::read(&output).unwrap(), b"existing destination");
+    assert_eq!(fs::read_dir(scratch.path()).unwrap().count(), 3);
+}