Add document-container crate (#4191)

* Add document-container crate: container backends and archive codecs

* Address PR review: path/prefix split, safe size casts, OPFS stream aborts

* Address PR review round 2: mmap read check, UTF8 entry names, prefix normalization

* Make MmappedBytes::new fallible so mmap reads can't silently degrade

* Address PR review round 3: backend contract uniformity (symlinks, remove, list)

* Apply symlink-component check to FolderBackend listing paths

* Address PR review: idempotent OPFS delete logging, tar default-features, shared entry-size cap

* Fix validate_path doc: dotfiles pass, CurDir/ParentDir rejected

* Omit symlink entries from FolderBackend listings for consistency with resolve

* Preserve zip I/O errors and reject non-canonical paths in validate_path

* Extend archive apis to return the archive writer

* Rename document/document-container directory to document/container

* Add archive format sniffing and deserialize_auto

* Drop temporal hedge from checked_entry_size comment

* Tighten verbose doc comments in document-container

* Coalesce consecutive same-path OPFS appends to avoid O(n^2) file copies

* Review

* Update document-container for the deserialize/store_non_blocking renames

---------

Co-authored-by: Timon <me@timon.zip>

Cargo.lock

@@ -1242,6 +1242,24 @@ version = "1.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c2db04e74f0a9a93103b50e90b96024c9b2bdca8bce6a632ec71b88736d3d359"
 
+[[package]]
+name = "document-container"
+version = "0.0.0"
+dependencies = [
+ "futures",
+ "js-sys",
+ "log",
+ "lzma-rust2",
+ "mmap-io",
+ "tar",
+ "tempfile",
+ "thiserror 2.0.18",
+ "wasm-bindgen",
+ "wasm-bindgen-futures",
+ "web-sys",
+ "zip",
+]
+
 [[package]]
 name = "document-features"
 version = "0.2.11"
@@ -1503,6 +1521,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4a3d7db9596fecd151c5f638c0ee5d5bd487b6e0ea232e5dc96d5250f6f94b1d"
 dependencies = [
  "crc32fast",
+ "libz-rs-sys",
  "miniz_oxide",
 ]
 
@@ -3005,6 +3024,15 @@ dependencies = [
  "redox_syscall",
 ]
 
+[[package]]
+name = "libz-rs-sys"
+version = "0.5.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c10501e7805cee23da17c7790e59df2870c0d4043ec6d03f67d31e2b53e77415"
+dependencies = [
+ "zlib-rs",
+]
+
 [[package]]
 name = "linebender_resource_handle"
 version = "0.1.1"
@@ -5495,6 +5523,19 @@ dependencies = [
  "xattr",
 ]
 
+[[package]]
+name = "tempfile"
+version = "3.23.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2d31c77bdf42a745371d260a26ca7163f1e0924b64afa0b688e61b5a9fa02f16"
+dependencies = [
+ "fastrand",
+ "getrandom 0.3.3",
+ "once_cell",
+ "rustix",
+ "windows-sys 0.61.2",
+]
+
 [[package]]
 name = "tendril"
 version = "0.4.3"
@@ -7685,11 +7726,18 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2d04a6b5381502aa6087c94c669499eb1602eb9c5e8198e534de571f7154809b"
 dependencies = [
  "crc32fast",
+ "flate2",
  "indexmap",
  "memchr",
  "typed-path",
 ]
 
+[[package]]
+name = "zlib-rs"
+version = "0.5.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "40990edd51aae2c2b6907af74ffb635029d5788228222c4bb811e9351c0caad3"
+
 [[package]]
 name = "zune-core"
 version = "0.4.12"

Cargo.toml

@@ -7,6 +7,7 @@ members = [
 	"desktop/platform/linux",
 	"desktop/platform/mac",
 	"desktop/platform/win",
+	"document/container",
 	"editor",
 	"frontend/wrapper",
 	"libraries/dyn-any",

document/container/Cargo.toml

@@ -0,0 +1,48 @@
+[package]
+name = "document-container"
+description = "Container abstraction for the on-disk side of the .gdd document format"
+edition.workspace = true
+version.workspace = true
+license.workspace = true
+authors.workspace = true
+
+[features]
+default = []
+zip = ["dep:zip"]
+xz = ["dep:lzma-rust2", "dep:tar"]
+
+[dependencies]
+thiserror = "2.0"
+log = { workspace = true }
+zip = { workspace = true, optional = true, features = ["deflate-flate2-zlib-rs"],  default-features = false}
+lzma-rust2 = { workspace = true, optional = true }
+tar = { version = "0.4", optional = true, default-features = false }
+
+[target.'cfg(not(target_family = "wasm"))'.dependencies]
+mmap-io = { workspace = true }
+
+[target.'cfg(target_family = "wasm")'.dependencies]
+web-sys = { workspace = true, features = [
+	"Navigator",
+	"DomException",
+	"Window",
+	"StorageManager",
+	"FileSystemCreateWritableOptions",
+	"FileSystemDirectoryHandle",
+	"FileSystemFileHandle",
+	"FileSystemGetFileOptions",
+	"FileSystemGetDirectoryOptions",
+	"FileSystemHandle",
+	"FileSystemHandleKind",
+	"FileSystemWritableFileStream",
+	"WritableStream",
+	"Blob",
+] }
+js-sys = { workspace = true }
+wasm-bindgen = { workspace = true }
+wasm-bindgen-futures = { workspace = true }
+futures = { workspace = true }
+
+[dev-dependencies]
+tempfile = "3"
+futures = { workspace = true }

document/container/src/archive.rs

@@ -0,0 +1,97 @@
+//! Archive codecs (zip, xz).
+//!
+//! Each codec streams entries in both directions: writers wrap an `io::Write` sink, and
+//! `deserialize` reads from any `io::Read + Seek` source and streams entries into any [`Container`].
+
+#[cfg(any(feature = "zip", feature = "xz"))]
+use crate::ContainerError;
+use crate::{Container, Result};
+use std::io::{Read, Seek, Write};
+
+/// Hard cap on the total decompressed size a codec will materialize from one archive.
+/// Defends against decompression bombs at the cost of refusing legitimately huge archives.
+#[cfg(any(feature = "zip", feature = "xz"))]
+pub(crate) const MAX_DECOMPRESSED_SIZE: u64 = 4 * 1024 * 1024 * 1024; // 4GB
+
+/// Fold one entry's declared `size` into the running `total` and return it as a `usize` for `write_sized`.
+/// Both codecs route entries through here so the decompression-bomb cap and 32-bit-safe conversion live in
+/// one place. `write_sized` pre-allocates the declared size, so an over-large one is rejected before that.
+#[cfg(any(feature = "zip", feature = "xz"))]
+pub(crate) fn checked_entry_size(total: &mut u64, size: u64) -> Result<usize> {
+	*total = total.saturating_add(size);
+	if *total > MAX_DECOMPRESSED_SIZE {
+		return Err(ContainerError::SizeLimitExceeded {
+			declared: *total,
+			limit: MAX_DECOMPRESSED_SIZE,
+		});
+	}
+
+	// `usize` is 32-bit on wasm, so convert fallibly to rule out a silent truncation into a smaller allocation.
+	usize::try_from(size).map_err(|_| ContainerError::SizeLimitExceeded {
+		declared: size,
+		limit: usize::MAX as u64,
+	})
+}
+
+#[cfg(feature = "zip")]
+mod zip;
+#[cfg(feature = "zip")]
+pub use zip::{Zip, ZipWriter};
+
+#[cfg(feature = "xz")]
+mod xz;
+#[cfg(feature = "xz")]
+pub use xz::{Xz, XzWriter};
+
+/// Streaming archive codec. The associated `Writer` type wraps a `Write + Seek` sink (zip needs
+/// `Seek` for the central directory; xz doesn't but `Seek` is free on file-like sinks) and
+/// accepts entries one at a time. `finish` flushes the codec's trailer and consumes the wrapper.
+pub trait Archive {
+	type Writer<W: Write + Seek>: ArchiveWriter
+	where
+		W: Write + Seek;
+
+	fn writer<W: Write + Seek>(output: W) -> Result<Self::Writer<W>>;
+
+	/// Read entries from `source` and write each into `dest`, streaming so neither the full
+	/// archive nor the full container ever sits in memory at once.
+	fn open<R: Read + Seek, C: Container>(source: R, dest: &mut C) -> Result<()>;
+}
+
+pub trait ArchiveWriter {
+	fn write_entry(&mut self, path: &str, bytes: &[u8]) -> Result<()>;
+	fn finish(self) -> Result<()>;
+}
+
+/// Archive container formats distinguishable by their leading magic bytes.
+#[derive(Copy, Clone, Debug, PartialEq, Eq)]
+pub enum ArchiveFormat {
+	Xz,
+	Zip,
+}
+
+impl ArchiveFormat {
+	/// Sniff the format from the leading magic bytes: xz streams start with `FD 37 7A 58 5A 00`,
+	/// zip archives with `50 4B 03 04` (`PK\x03\x04`). Returns `None` for anything else.
+	pub fn detect(bytes: &[u8]) -> Option<Self> {
+		if bytes.starts_with(&[0xFD, 0x37, 0x7A, 0x58, 0x5A, 0x00]) {
+			Some(Self::Xz)
+		} else if bytes.starts_with(&[0x50, 0x4B, 0x03, 0x04]) {
+			Some(Self::Zip)
+		} else {
+			None
+		}
+	}
+}
+
+/// Deserialize an archive into `dest`, auto-detecting the format from `bytes`' magic header.
+/// Errors if the bytes are neither a recognized xz nor zip archive.
+#[cfg(all(feature = "xz", feature = "zip"))]
+pub fn open_auto<C: Container>(bytes: &[u8], dest: &mut C) -> Result<()> {
+	let source = std::io::Cursor::new(bytes);
+	match ArchiveFormat::detect(bytes) {
+		Some(ArchiveFormat::Xz) => Xz::open(source, dest),
+		Some(ArchiveFormat::Zip) => Zip::open(source, dest),
+		None => Err(ContainerError::Codec("unrecognized archive format (not xz or zip)".into())),
+	}
+}

document/container/src/archive/xz.rs

@@ -0,0 +1,97 @@
+//! Xz-compressed tarball archive codec.
+
+use crate::archive::{Archive, ArchiveWriter, MAX_DECOMPRESSED_SIZE, checked_entry_size};
+use crate::{Container, ContainerError, Result, validate_path};
+use lzma_rust2::{XzOptions, XzReader, XzWriter as InnerXzWriter};
+use std::io::{Read, Seek, Write};
+
+pub struct Xz;
+
+/// xz-tar writer. Held as an `Option` so `finish` can take ownership and unwind the layered
+/// writers in the right order: drop the tar builder first to flush its trailer, then finish xz.
+pub struct XzWriter<W: Write + Seek> {
+	tar: Option<tar::Builder<InnerXzWriter<W>>>,
+}
+
+impl Archive for Xz {
+	type Writer<W: Write + Seek> = XzWriter<W>;
+
+	fn writer<W: Write + Seek>(output: W) -> Result<Self::Writer<W>> {
+		let xz_writer = InnerXzWriter::new(output, XzOptions::default()).map_err(lzma_err)?;
+		Ok(XzWriter {
+			tar: Some(tar::Builder::new(xz_writer)),
+		})
+	}
+
+	fn open<R: Read + Seek, C: Container>(source: R, dest: &mut C) -> Result<()> {
+		// `take` bounds how many bytes we decompress from the xz stream, but each tar entry's declared
+		// size is fed to `write_sized`, which pre-allocates from it before reading. Cap the cumulative
+		// declared size too so a header claiming a huge size can't trigger a giant allocation up front.
+		let xz_reader = XzReader::new(source, false);
+		let bounded = xz_reader.take(MAX_DECOMPRESSED_SIZE);
+
+		let mut tar_reader = tar::Archive::new(bounded);
+		let mut total_size = 0u64;
+
+		for entry in tar_reader.entries()? {
+			let mut entry = entry?;
+			if entry.header().entry_type() != tar::EntryType::Regular {
+				continue;
+			}
+			// Reject non-UTF8 entry names rather than lossily rewriting them, so the path we store matches
+			// the archive exactly.
+			let path = entry.path()?;
+			let path = path.to_str().ok_or_else(|| ContainerError::Codec(format!("tar: non-UTF8 entry name {path:?}")))?.to_owned();
+			validate_path(&path)?;
+
+			let size = checked_entry_size(&mut total_size, entry.size())?;
+
+			dest.write_sized(&path, size, &mut |buffer| {
+				entry.read_exact(buffer).map_err(ContainerError::Io)?;
+				Ok(())
+			})?;
+		}
+
+		Ok(())
+	}
+}
+
+impl<W: Write + Seek> ArchiveWriter for XzWriter<W> {
+	fn write_entry(&mut self, path: &str, bytes: &[u8]) -> Result<()> {
+		validate_path(path)?;
+		let tar = self.tar.as_mut().ok_or_else(|| ContainerError::Codec("XzWriter already finished".into()))?;
+		let mut header = tar::Header::new_gnu();
+		header.set_path(path).map_err(|error| ContainerError::Codec(format!("tar: invalid path {path}: {error}")))?;
+		header.set_size(bytes.len() as u64);
+		header.set_mode(0o644);
+		header.set_cksum();
+		tar.append(&header, bytes)?;
+		Ok(())
+	}
+
+	fn finish(mut self) -> Result<()> {
+		self.finish_inner()?;
+		Ok(())
+	}
+}
+
+impl<W: Write + Seek> XzWriter<W> {
+	/// Finish the archive and return the underlying sink, for in-memory archives where the caller
+	/// wants the written bytes (e.g. `Cursor<Vec<u8>>`) back.
+	pub fn finish_into(mut self) -> Result<W> {
+		self.finish_inner()
+	}
+
+	/// Unwind the layered writers in order (flush the tar trailer, then finish xz) and hand back the
+	/// innermost sink. Shared by `finish` and `finish_into`.
+	fn finish_inner(&mut self) -> Result<W> {
+		let mut tar = self.tar.take().ok_or_else(|| ContainerError::Codec("XzWriter already finished".into()))?;
+		tar.finish()?;
+		let xz_writer = tar.into_inner()?;
+		xz_writer.finish().map_err(lzma_err)
+	}
+}
+
+fn lzma_err(error: std::io::Error) -> ContainerError {
+	ContainerError::Codec(format!("lzma: {error}"))
+}

document/container/src/archive/zip.rs

@@ -0,0 +1,83 @@
+//! Zip archive codec.
+
+use crate::archive::{Archive, ArchiveWriter, checked_entry_size};
+use crate::{Container, ContainerError, Result, validate_path};
+use std::io::{Read, Seek, Write};
+
+use zip::ZipArchive;
+use zip::write::{SimpleFileOptions, ZipWriter as InnerZipWriter};
+
+pub struct Zip;
+
+pub struct ZipWriter<W: Write + Seek> {
+	inner: InnerZipWriter<W>,
+	options: SimpleFileOptions,
+}
+
+impl Archive for Zip {
+	type Writer<W: Write + Seek> = ZipWriter<W>;
+
+	fn writer<W: Write + Seek>(output: W) -> Result<Self::Writer<W>> {
+		Ok(ZipWriter {
+			inner: InnerZipWriter::new(output),
+			options: SimpleFileOptions::default().compression_method(zip::CompressionMethod::Deflated),
+		})
+	}
+
+	fn open<R: Read + Seek, C: Container>(source: R, dest: &mut C) -> Result<()> {
+		let mut archive = ZipArchive::new(source).map_err(zip_err)?;
+
+		// Zip headers declare each entry's uncompressed size up front; `checked_entry_size` caps the running
+		// total so a malicious archive can't exhaust memory or disk before any bytes are read.
+		let mut total_size = 0u64;
+
+		for index in 0..archive.len() {
+			let mut entry = archive.by_index(index).map_err(zip_err)?;
+			if !entry.is_file() {
+				continue;
+			}
+			let name = entry.name().to_string();
+			validate_path(&name)?;
+
+			let size = checked_entry_size(&mut total_size, entry.size())?;
+
+			dest.write_sized(&name, size, &mut |buffer| {
+				entry.read_exact(buffer).map_err(ContainerError::Io)?;
+				Ok(())
+			})?;
+		}
+
+		Ok(())
+	}
+}
+
+impl<W: Write + Seek> ArchiveWriter for ZipWriter<W> {
+	fn write_entry(&mut self, path: &str, bytes: &[u8]) -> Result<()> {
+		validate_path(path)?;
+		self.inner.start_file(path, self.options).map_err(zip_err)?;
+		self.inner.write_all(bytes)?;
+		Ok(())
+	}
+
+	fn finish(self) -> Result<()> {
+		self.inner.finish().map_err(zip_err)?;
+		Ok(())
+	}
+}
+
+impl<W: Write + Seek> ZipWriter<W> {
+	/// Finish the archive and return the underlying sink, for in-memory archives where the caller
+	/// wants the written bytes (e.g. `Cursor<Vec<u8>>`) back.
+	pub fn finish_into(self) -> Result<W> {
+		self.inner.finish().map_err(zip_err)
+	}
+}
+
+fn zip_err(error: zip::result::ZipError) -> ContainerError {
+	// Preserve a real I/O failure (disk full, etc.) as a structured `Io` so callers can tell it apart from a
+	// corrupt-archive `Codec` error; only the genuinely archive-format errors collapse into `Codec`.
+	match error {
+		zip::result::ZipError::Io(io) => ContainerError::Io(io),
+		other => ContainerError::Codec(format!("zip: {other}")),
+	}
+}

document/container/src/backends.rs

@@ -0,0 +1,9 @@
+//! Container backend implementations.
+
+pub mod memory;
+
+#[cfg(not(target_family = "wasm"))]
+pub mod folder;
+
+#[cfg(target_family = "wasm")]
+pub mod opfs;

document/container/src/backends/folder.rs

@@ -0,0 +1,191 @@
+//! Loose-folder backend.
+
+use crate::{ByteHolder, Container, ContainerError, MmappedBytes, Result, validate_path, validate_prefix};
+use mmap_io::mmap::{MemoryMappedFile, MmapMode};
+use std::fs::{self, OpenOptions};
+use std::io::Write;
+use std::path::{Path, PathBuf};
+
+pub struct FolderBackend {
+	root: PathBuf,
+}
+
+impl FolderBackend {
+	/// Open an existing folder. Errors if `root` is not a directory.
+	pub fn open(root: impl Into<PathBuf>) -> Result<Self> {
+		let root = root.into();
+		if !root.is_dir() {
+			return Err(ContainerError::NotFound(root.display().to_string()));
+		}
+		Ok(Self { root })
+	}
+
+	/// Create the folder if it does not exist, then open it.
+	pub fn create(root: impl Into<PathBuf>) -> Result<Self> {
+		let root = root.into();
+		fs::create_dir_all(&root)?;
+		Ok(Self { root })
+	}
+
+	pub fn root(&self) -> &std::path::Path {
+		&self.root
+	}
+
+	fn resolve(&self, path: &str) -> Result<PathBuf> {
+		validate_path(path)?;
+		self.reject_symlinked_components(path)?;
+		Ok(self.root.join(path))
+	}
+
+	/// Reject any existing component along `root/relative` that is a symlink. `validate_path`/`validate_prefix`
+	/// block `..` and absolute paths, but a symlink stored under the root could still point outside it, so every
+	/// path that gets joined onto the root must pass through here before it is opened or traversed.
+	fn reject_symlinked_components(&self, relative: &str) -> Result<()> {
+		let mut partial = self.root.clone();
+		for component in Path::new(relative).components() {
+			partial.push(component);
+			if let Ok(metadata) = fs::symlink_metadata(&partial)
+				&& metadata.file_type().is_symlink()
+			{
+				return Err(ContainerError::InvalidPath(relative.to_string()));
+			}
+		}
+		Ok(())
+	}
+
+	fn list_filtered(&self, prefix: &str, want_files: bool) -> Result<Vec<String>> {
+		validate_prefix(prefix)?;
+		let base = if prefix.is_empty() || prefix == "." {
+			self.root.clone()
+		} else {
+			self.reject_symlinked_components(prefix)?;
+			self.root.join(prefix)
+		};
+
+		// A missing prefix has no entries; a prefix that names a file is a misuse.
+		if base.is_file() {
+			return Err(ContainerError::NotADirectory(prefix.to_string()));
+		}
+		if !base.is_dir() {
+			return Ok(Vec::new());
+		}
+
+		let mut results = Vec::new();
+		for entry in fs::read_dir(&base)? {
+			let entry = entry?;
+
+			// `file_type` does not follow symlinks, unlike `is_file`/`is_dir`. Skip symlink entries so a
+			// listing never advertises a path that `resolve` would then reject as a container escape.
+			let Ok(file_type) = entry.file_type() else { continue };
+			let matches = if want_files { file_type.is_file() } else { file_type.is_dir() };
+			if !matches {
+				continue;
+			}
+
+			let path = entry.path();
+			let relative = path.strip_prefix(&self.root).map_err(|_| ContainerError::Backend("path escaped root".into()))?;
+			results.push(relative.to_string_lossy().replace('\\', "/"));
+		}
+		Ok(results)
+	}
+}
+
+impl Container for FolderBackend {
+	fn read(&self, path: &str) -> Result<ByteHolder> {
+		let full = self.resolve(path)?;
+		if !full.is_file() {
+			return Err(ContainerError::NotFound(path.to_string()));
+		}
+
+		// Mmapping a zero-length file is platform-dependent and often fails, so serve empty files as owned
+		// bytes and reserve mmap for files that actually have content.
+		if fs::metadata(&full).map(|metadata| metadata.len() == 0).unwrap_or(false) {
+			return Ok(ByteHolder::Owned(Vec::new()));
+		}
+
+		Ok(ByteHolder::Mmapped(MmappedBytes::new(open_mmap(&full)?)?))
+	}
+
+	fn write(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		let full = self.resolve(path)?;
+		if let Some(parent) = full.parent() {
+			fs::create_dir_all(parent)?;
+		}
+		fs::write(&full, bytes)?;
+		Ok(())
+	}
+
+	fn append(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		let full = self.resolve(path)?;
+		if let Some(parent) = full.parent() {
+			fs::create_dir_all(parent)?;
+		}
+		let mut file = OpenOptions::new().create(true).append(true).open(&full)?;
+		file.write_all(bytes)?;
+		Ok(())
+	}
+
+	fn write_sized(&self, path: &str, size: usize, fill: &mut dyn FnMut(&mut [u8]) -> Result<()>) -> Result<()> {
+		if size == 0 {
+			return self.write(path, &[]);
+		}
+
+		let full = self.resolve(path)?;
+		if let Some(parent) = full.parent() {
+			fs::create_dir_all(parent)?;
+		}
+
+		let file = MemoryMappedFile::create_rw(&full, size as u64).map_err(|error| ContainerError::Backend(format!("create_rw {full:?} failed: {error}")))?;
+
+		let result = {
+			let mut slice = file
+				.as_slice_mut(0, size as u64)
+				.map_err(|error| ContainerError::Backend(format!("as_slice_mut {full:?} failed: {error}")))?;
+			fill(slice.as_mut())?;
+			drop(slice);
+			file.flush().map_err(|error| ContainerError::Backend(format!("flush {full:?} failed: {error}")))
+		};
+
+		// `create_rw` materializes the full-size file before `fill` runs, so remove it on failure rather
+		// than leave a zeroed or half-written remnant.
+		if result.is_err() {
+			drop(file);
+			let _ = fs::remove_file(&full);
+		}
+		result
+	}
+
+	fn list(&self, prefix: &str) -> Result<Vec<String>> {
+		self.list_filtered(prefix, true)
+	}
+
+	fn list_dirs(&self, prefix: &str) -> Result<Vec<String>> {
+		self.list_filtered(prefix, false)
+	}
+
+	fn exists(&self, path: &str) -> bool {
+		match self.resolve(path) {
+			Ok(full) => full.is_file(),
+			Err(_) => false,
+		}
+	}
+
+	fn remove(&self, path: &str) -> Result<()> {
+		let full = self.resolve(path)?;
+		// Idempotent: a missing file is not an error. Directories are left alone.
+		if full.is_file() {
+			fs::remove_file(full)?;
+		}
+		// TODO: decide if we should remove empty parent directories
+		Ok(())
+	}
+}
+
+/// Open a memory-mapped read-only view of `path`, trying huge pages first. Callers must ensure `path`
+/// is non-empty, since mmapping a zero-length file is platform-dependent.
+fn open_mmap(path: &Path) -> Result<MemoryMappedFile> {
+	match MemoryMappedFile::builder(path).mode(MmapMode::ReadOnly).huge_pages(true).open() {
+		Ok(file) => Ok(file),
+		Err(_) => MemoryMappedFile::open_ro(path).map_err(|error| ContainerError::Backend(format!("mmap of {path:?} failed: {error}"))),
+	}
+}

document/container/src/backends/memory.rs

@@ -0,0 +1,89 @@
+//! In-memory backend. Useful for tests and as the deserialize target for archive codecs.
+
+use crate::{ByteHolder, Container, ContainerError, Result, validate_path, validate_prefix, with_trailing_slash};
+use std::collections::{HashMap, HashSet};
+use std::sync::Mutex;
+
+#[derive(Default)]
+pub struct MemoryBackend {
+	files: Mutex<HashMap<String, Vec<u8>>>,
+}
+
+impl MemoryBackend {
+	pub fn new() -> Self {
+		Self::default()
+	}
+}
+
+impl Container for MemoryBackend {
+	fn read(&self, path: &str) -> Result<ByteHolder> {
+		validate_path(path)?;
+		self.files
+			.lock()
+			.unwrap()
+			.get(path)
+			.map(|bytes| ByteHolder::Owned(bytes.clone()))
+			.ok_or_else(|| ContainerError::NotFound(path.to_string()))
+	}
+
+	fn write(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		validate_path(path)?;
+		self.files.lock().unwrap().insert(path.to_string(), bytes.to_vec());
+		Ok(())
+	}
+
+	fn append(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		validate_path(path)?;
+		self.files.lock().unwrap().entry(path.to_string()).or_default().extend_from_slice(bytes);
+		Ok(())
+	}
+
+	fn list(&self, prefix: &str) -> Result<Vec<String>> {
+		validate_prefix(prefix)?;
+		let files = self.files.lock().unwrap();
+		// A prefix that is itself a stored key names a file, not a directory.
+		if files.contains_key(prefix) {
+			return Err(ContainerError::NotADirectory(prefix.to_string()));
+		}
+
+		let normalized = with_trailing_slash(prefix);
+		let results = files.keys().filter(|path| path.starts_with(&normalized) && !path[normalized.len()..].contains('/')).cloned().collect();
+		Ok(results)
+	}
+
+	fn list_dirs(&self, prefix: &str) -> Result<Vec<String>> {
+		validate_prefix(prefix)?;
+		let files = self.files.lock().unwrap();
+		if files.contains_key(prefix) {
+			return Err(ContainerError::NotADirectory(prefix.to_string()));
+		}
+
+		let normalized = with_trailing_slash(prefix);
+		let mut seen = HashSet::new();
+		let mut results = Vec::new();
+		for path in files.keys() {
+			if !path.starts_with(&normalized) {
+				continue;
+			}
+			let remainder = &path[normalized.len()..];
+			if let Some((segment, _)) = remainder.split_once('/') {
+				let dir = format!("{normalized}{segment}");
+				if seen.insert(dir.clone()) {
+					results.push(dir);
+				}
+			}
+		}
+		Ok(results)
+	}
+
+	fn exists(&self, path: &str) -> bool {
+		validate_path(path).is_ok() && self.files.lock().unwrap().contains_key(path)
+	}
+
+	fn remove(&self, path: &str) -> Result<()> {
+		validate_path(path)?;
+		// Idempotent: removing a missing path is not an error.
+		self.files.lock().unwrap().remove(path);
+		Ok(())
+	}
+}

document/container/src/backends/opfs.rs

@@ -0,0 +1,475 @@
+//! OPFS (Origin Private File System) backend for browser wasm.
+
+use crate::{AsyncContainer, ByteHolder, ContainerError, Result, validate_path, validate_prefix, with_trailing_slash};
+use futures::channel::oneshot;
+use js_sys::Uint8Array;
+use std::collections::{HashSet, VecDeque};
+use std::sync::{Arc, Mutex};
+use wasm_bindgen::JsCast;
+use wasm_bindgen::prelude::*;
+use wasm_bindgen_futures::{JsFuture, spawn_local};
+use web_sys::{
+	Blob, DomException, FileSystemCreateWritableOptions, FileSystemDirectoryHandle, FileSystemFileHandle, FileSystemGetDirectoryOptions, FileSystemGetFileOptions, FileSystemWritableFileStream,
+	WritableStream,
+};
+
+enum Mutation {
+	Write {
+		path: String,
+		bytes: Vec<u8>,
+	},
+	Append {
+		path: String,
+		bytes: Vec<u8>,
+	},
+	Delete {
+		path: String,
+	},
+	/// In-band flush barrier. The worker fulfills the sender once it dequeues this, which (by FIFO
+	/// order) signals that every mutation enqueued before it has been applied to disk.
+	Barrier(oneshot::Sender<()>),
+}
+
+struct Inner {
+	directory: FileSystemDirectoryHandle,
+	/// Paths believed to be on disk, letting `exists_non_blocking` answer without OPFS's missing sync API.
+	/// Optimistic: a path is inserted when its write is queued and kept even if that write later fails, so it
+	/// can briefly over-report. Rebuilt from real disk state on the next `open`.
+	on_disk: HashSet<String>,
+	queue: VecDeque<Mutation>,
+	worker_active: bool,
+}
+
+pub struct OpfsBackend {
+	inner: Arc<Mutex<Inner>>,
+}
+
+// Safety: only built for browser wasm where JS handles never leave the main thread.
+unsafe impl Send for OpfsBackend {}
+unsafe impl Sync for OpfsBackend {}
+
+impl OpfsBackend {
+	/// Open (or create) `directory_name` under the OPFS root.
+	pub async fn open(directory_name: &str) -> Result<Self> {
+		let directory = open_directory(directory_name).await.map_err(js_err)?;
+		let on_disk = enumerate_paths(&directory, "").await.map_err(js_err)?;
+		Ok(Self {
+			inner: Arc::new(Mutex::new(Inner {
+				directory,
+				on_disk,
+				queue: VecDeque::new(),
+				worker_active: false,
+			})),
+		})
+	}
+
+	fn directory(&self) -> FileSystemDirectoryHandle {
+		self.inner.lock().unwrap().directory.clone()
+	}
+
+	/// Wait until all queued non-blocking mutations have hit disk, so awaited reads observe them. Plants a
+	/// barrier at the back of the queue and awaits the worker reaching it; by FIFO order, everything ahead is
+	/// then applied. Draining stays the worker's job so it remains the sole mutator.
+	async fn flush(&self) {
+		let (sender, receiver) = oneshot::channel();
+		{
+			let mut guard = self.inner.lock().unwrap();
+			guard.queue.push_back(Mutation::Barrier(sender));
+			kick_worker(&self.inner, &mut guard);
+		}
+		// The sender is only dropped without sending if the worker is torn down mid-drain; either way
+		// there is nothing left to wait for, so a receive error is treated as "already flushed".
+		let _ = receiver.await;
+	}
+}
+
+impl AsyncContainer for OpfsBackend {
+	async fn read(&self, path: &str) -> Result<ByteHolder> {
+		validate_path(path)?;
+
+		// Non-blocking writes only land on disk once the queue drains, so flush first and then treat
+		// disk as authoritative. Draining (rather than folding the queue against an awaited base read)
+		// avoids racing the background worker, which could otherwise double-apply a queued append.
+		self.flush().await;
+
+		let bytes = read_file(&self.directory(), path).await.map_err(js_err)?;
+		Ok(ByteHolder::Owned(bytes))
+	}
+
+	async fn write(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		validate_path(path)?;
+		// Flush first so this awaited write lands after any non-blocking mutation already queued for the
+		// same path, then apply directly so the real disk error still propagates to the caller.
+		self.flush().await;
+		write_file(&self.directory(), path, bytes).await.map_err(js_err)?;
+		self.inner.lock().unwrap().on_disk.insert(path.to_string());
+		Ok(())
+	}
+
+	async fn append(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		validate_path(path)?;
+		self.flush().await;
+		append_file(&self.directory(), path, bytes).await.map_err(js_err)?;
+		self.inner.lock().unwrap().on_disk.insert(path.to_string());
+		Ok(())
+	}
+
+	async fn write_sized(&self, path: &str, size: usize, fill: &mut dyn FnMut(&mut [u8]) -> Result<()>) -> Result<()> {
+		validate_path(path)?;
+		let mut buffer = vec![0; size];
+		fill(&mut buffer)?;
+		self.flush().await;
+		write_file(&self.directory(), path, &buffer).await.map_err(js_err)?;
+		self.inner.lock().unwrap().on_disk.insert(path.to_string());
+		Ok(())
+	}
+
+	async fn list(&self, prefix: &str) -> Result<Vec<String>> {
+		validate_prefix(prefix)?;
+		self.flush().await;
+		list_entries(&self.directory(), prefix, EntryKind::File).await.map_err(|error| list_error(prefix, error))
+	}
+
+	async fn list_dirs(&self, prefix: &str) -> Result<Vec<String>> {
+		validate_prefix(prefix)?;
+		self.flush().await;
+		list_entries(&self.directory(), prefix, EntryKind::Directory).await.map_err(|error| list_error(prefix, error))
+	}
+
+	async fn exists(&self, path: &str) -> bool {
+		if validate_path(path).is_err() {
+			return false;
+		}
+		self.flush().await;
+		file_exists(&self.directory(), path).await
+	}
+
+	async fn remove(&self, path: &str) -> Result<()> {
+		validate_path(path)?;
+		self.flush().await;
+		// Idempotent: a missing entry is not an error, matching the other backends.
+		if let Err(error) = remove_file(&self.directory(), path).await
+			&& !is_not_found(&error)
+		{
+			return Err(js_err(error));
+		}
+		self.inner.lock().unwrap().on_disk.remove(path);
+		Ok(())
+	}
+
+	fn write_non_blocking(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		validate_path(path)?;
+		let mut guard = self.inner.lock().unwrap();
+		guard.on_disk.insert(path.to_string());
+		guard.queue.push_back(Mutation::Write {
+			path: path.to_string(),
+			bytes: bytes.to_vec(),
+		});
+		kick_worker(&self.inner, &mut guard);
+		Ok(())
+	}
+
+	fn append_non_blocking(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		validate_path(path)?;
+		let mut guard = self.inner.lock().unwrap();
+		guard.on_disk.insert(path.to_string());
+		guard.queue.push_back(Mutation::Append {
+			path: path.to_string(),
+			bytes: bytes.to_vec(),
+		});
+		kick_worker(&self.inner, &mut guard);
+		Ok(())
+	}
+
+	fn remove_non_blocking(&self, path: &str) -> Result<()> {
+		validate_path(path)?;
+		let mut guard = self.inner.lock().unwrap();
+		guard.on_disk.remove(path);
+		guard.queue.push_back(Mutation::Delete { path: path.to_string() });
+		kick_worker(&self.inner, &mut guard);
+		Ok(())
+	}
+
+	fn exists_non_blocking(&self, path: &str) -> bool {
+		validate_path(path).is_ok() && self.inner.lock().unwrap().on_disk.contains(path)
+	}
+}
+
+fn kick_worker(inner: &Arc<Mutex<Inner>>, guard: &mut Inner) {
+	if guard.worker_active {
+		return;
+	}
+	guard.worker_active = true;
+	let inner = inner.clone();
+	spawn_local(drain_queue(inner));
+}
+
+/// Apply every queued mutation to disk, in FIFO order, until the queue is empty, then mark the worker
+/// idle. Spawned once via [`kick_worker`] and runs as the sole mutator; the awaited read paths wait on
+/// a [`Mutation::Barrier`] rather than draining themselves, so there is never more than one drainer.
+///
+/// Consecutive appends to the same path are coalesced into one combined append. Each `createWritable`
+/// on OPFS copies the whole existing file, so appending N frames one at a time is O(N^2) in bytes
+/// copied (every per-op autosave queues one append per hot op). Concatenating a run into a single
+/// append collapses that to one file copy, while staying byte-identical to applying them in order.
+async fn drain_queue(inner: Arc<Mutex<Inner>>) {
+	loop {
+		// Pop the front mutation, coalescing a run of same-path appends behind it into one batch.
+		let (directory, mutation) = {
+			let mut guard = inner.lock().unwrap();
+			let Some(mut mutation) = guard.queue.pop_front() else {
+				guard.worker_active = false;
+				return;
+			};
+
+			if let Mutation::Append { path, bytes } = &mut mutation {
+				while let Some(Mutation::Append { path: next_path, .. }) = guard.queue.front()
+					&& next_path == path
+				{
+					let Some(Mutation::Append { bytes: next_bytes, .. }) = guard.queue.pop_front() else {
+						unreachable!()
+					};
+					bytes.extend_from_slice(&next_bytes);
+				}
+			}
+
+			(guard.directory.clone(), mutation)
+		};
+
+		match mutation {
+			Mutation::Write { path, bytes } => {
+				if let Err(error) = write_file(&directory, &path, &bytes).await {
+					log::error!("OPFS background write for {path} failed: {error:?}");
+				}
+			}
+			Mutation::Append { path, bytes } => {
+				if let Err(error) = append_file(&directory, &path, &bytes).await {
+					log::error!("OPFS background append for {path} failed: {error:?}");
+				}
+			}
+			Mutation::Delete { path } => {
+				// Removal is idempotent, so a missing entry is the expected outcome of a redundant delete, not an error.
+				if let Err(error) = remove_file(&directory, &path).await
+					&& !is_not_found(&error)
+				{
+					log::error!("OPFS background delete for {path} failed: {error:?}");
+				}
+			}
+			// A receive error on the waiter side just means the reader gave up; nothing to apply.
+			Mutation::Barrier(sender) => {
+				let _ = sender.send(());
+			}
+		}
+	}
+}
+
+fn js_err(error: JsValue) -> ContainerError {
+	ContainerError::Backend(format!("{error:?}"))
+}
+
+/// Resolve `directory_path` (a `/`-separated relative path) under the OPFS root, creating each
+/// segment. OPFS rejects directory names containing `/`, so a multi-segment path like
+/// `documents/<id>` must be descended one segment at a time rather than passed whole.
+async fn open_directory(directory_path: &str) -> std::result::Result<FileSystemDirectoryHandle, JsValue> {
+	let storage = web_sys::window().ok_or_else(|| JsValue::from_str("no window"))?.navigator().storage();
+	let mut current: FileSystemDirectoryHandle = JsFuture::from(storage.get_directory()).await?.dyn_into()?;
+
+	for segment in directory_path.split('/').filter(|segment| !segment.is_empty()) {
+		let options = FileSystemGetDirectoryOptions::new();
+		options.set_create(true);
+		current = JsFuture::from(current.get_directory_handle_with_options(segment, &options)).await?.dyn_into()?;
+	}
+
+	Ok(current)
+}
+
+/// Descend the `/`-separated path against `root` and return the directory handle plus the final segment.
+async fn descend<'a>(root: &FileSystemDirectoryHandle, relative: &'a str, create_dirs: bool) -> std::result::Result<(FileSystemDirectoryHandle, &'a str), JsValue> {
+	let mut current = root.clone();
+	let mut segments = relative.split('/').filter(|s| !s.is_empty()).collect::<Vec<_>>();
+	let file = segments.pop().ok_or_else(|| JsValue::from_str("empty path"))?;
+
+	for segment in segments {
+		let options = FileSystemGetDirectoryOptions::new();
+		options.set_create(create_dirs);
+		current = JsFuture::from(current.get_directory_handle_with_options(segment, &options)).await?.dyn_into()?;
+	}
+	Ok((current, file))
+}
+
+async fn read_file(root: &FileSystemDirectoryHandle, path: &str) -> std::result::Result<Vec<u8>, JsValue> {
+	let (directory, name) = descend(root, path, false).await?;
+
+	let handle: FileSystemFileHandle = JsFuture::from(directory.get_file_handle(name)).await?.dyn_into()?;
+	let file_value = JsFuture::from(handle.get_file()).await?;
+	let blob: Blob = file_value.dyn_into()?;
+	let buffer = JsFuture::from(blob.array_buffer()).await?;
+	Ok(Uint8Array::new(&buffer).to_vec())
+}
+
+async fn write_file(root: &FileSystemDirectoryHandle, path: &str, bytes: &[u8]) -> std::result::Result<(), JsValue> {
+	let (directory, name) = descend(root, path, true).await?;
+
+	let options = FileSystemGetFileOptions::new();
+	options.set_create(true);
+
+	let handle: FileSystemFileHandle = JsFuture::from(directory.get_file_handle_with_options(name, &options)).await?.dyn_into()?;
+	let writable: FileSystemWritableFileStream = JsFuture::from(handle.create_writable()).await?.dyn_into()?;
+	let stream: WritableStream = writable.clone().unchecked_into();
+
+	// Wrap in an async block so any failure, including a synchronous JS throw from `write_with_js_u8_array`,
+	// aborts the stream before returning instead of leaving a dangling locked writable.
+	let write = async {
+		let array = Uint8Array::from(bytes);
+		JsFuture::from(writable.write_with_js_u8_array(&array)?).await?;
+		Ok::<(), JsValue>(())
+	}
+	.await;
+
+	if let Err(error) = write {
+		let _ = JsFuture::from(stream.abort()).await;
+		return Err(error);
+	}
+
+	JsFuture::from(stream.close()).await?;
+	Ok(())
+}
+
+async fn append_file(root: &FileSystemDirectoryHandle, path: &str, bytes: &[u8]) -> std::result::Result<(), JsValue> {
+	let (directory, name) = descend(root, path, true).await?;
+
+	let file_options = FileSystemGetFileOptions::new();
+	file_options.set_create(true);
+	let handle: FileSystemFileHandle = JsFuture::from(directory.get_file_handle_with_options(name, &file_options)).await?.dyn_into()?;
+
+	// Determine the current end-of-file so we can seek there before writing.
+	let file_value = JsFuture::from(handle.get_file()).await?;
+	let blob: Blob = file_value.dyn_into()?;
+	let offset = blob.size();
+
+	// `keepExistingData: true` preserves bytes outside the written range; without it OPFS truncates to the written length.
+	let writable_options = FileSystemCreateWritableOptions::new();
+	writable_options.set_keep_existing_data(true);
+	let writable: FileSystemWritableFileStream = JsFuture::from(handle.create_writable_with_options(&writable_options)).await?.dyn_into()?;
+	let stream: WritableStream = writable.clone().unchecked_into();
+
+	// Wrap in an async block so a synchronous JS throw from `seek_with_f64`/`write_with_js_u8_array`
+	// aborts the stream instead of bypassing the abort via `?`.
+	let seek_and_write = async {
+		JsFuture::from(writable.seek_with_f64(offset)?).await?;
+		let array = Uint8Array::from(bytes);
+		JsFuture::from(writable.write_with_js_u8_array(&array)?).await?;
+		Ok::<(), JsValue>(())
+	}
+	.await;
+
+	if let Err(error) = seek_and_write {
+		let _ = JsFuture::from(stream.abort()).await;
+		return Err(error);
+	}
+
+	JsFuture::from(stream.close()).await?;
+	Ok(())
+}
+
+async fn remove_file(root: &FileSystemDirectoryHandle, path: &str) -> std::result::Result<(), JsValue> {
+	let (directory, name) = descend(root, path, false).await?;
+	JsFuture::from(directory.remove_entry(name)).await?;
+	Ok(())
+}
+
+async fn file_exists(root: &FileSystemDirectoryHandle, path: &str) -> bool {
+	let Ok((directory, name)) = descend(root, path, false).await else {
+		return false;
+	};
+	JsFuture::from(directory.get_file_handle(name)).await.is_ok()
+}
+
+#[derive(Clone, Copy)]
+enum EntryKind {
+	File,
+	Directory,
+}
+
+async fn list_entries(root: &FileSystemDirectoryHandle, prefix: &str, want: EntryKind) -> std::result::Result<Vec<String>, JsValue> {
+	// `.` and the empty string both name the container root.
+	let prefix = if prefix == "." { "" } else { prefix };
+
+	let directory = if prefix.is_empty() {
+		root.clone()
+	} else {
+		let mut current = root.clone();
+		for segment in prefix.split('/').filter(|s| !s.is_empty()) {
+			let options = FileSystemGetDirectoryOptions::new();
+			options.set_create(false);
+			current = match JsFuture::from(current.get_directory_handle_with_options(segment, &options)).await {
+				Ok(value) => value.dyn_into()?,
+				Err(error) if is_not_found(&error) => return Ok(Vec::new()),
+				Err(error) => return Err(error),
+			};
+		}
+		current
+	};
+
+	let entries = directory.entries();
+	let iterator: js_sys::AsyncIterator = entries.unchecked_into();
+	let mut results = Vec::new();
+	let prefix_with_slash = with_trailing_slash(prefix);
+
+	let want_kind = match want {
+		EntryKind::File => web_sys::FileSystemHandleKind::File,
+		EntryKind::Directory => web_sys::FileSystemHandleKind::Directory,
+	};
+
+	loop {
+		let next: js_sys::IteratorNext = JsFuture::from(iterator.next()?).await?.unchecked_into();
+		if next.done() {
+			break;
+		}
+		let pair: js_sys::Array = next.value().unchecked_into();
+		let Some(name) = pair.get(0).as_string() else { continue };
+		let handle: web_sys::FileSystemHandle = pair.get(1).unchecked_into();
+		if handle.kind() != want_kind {
+			continue;
+		}
+		results.push(format!("{prefix_with_slash}{name}"));
+	}
+	Ok(results)
+}
+
+/// Walk every file under `prefix` (recursively) and collect their full container paths.
+/// Used at open time to populate the in-memory tracking set.
+async fn enumerate_paths(root: &FileSystemDirectoryHandle, prefix: &str) -> std::result::Result<HashSet<String>, JsValue> {
+	let mut paths = HashSet::new();
+	let mut to_visit = vec![prefix.to_string()];
+
+	while let Some(current_prefix) = to_visit.pop() {
+		for file in list_entries(root, &current_prefix, EntryKind::File).await? {
+			paths.insert(file);
+		}
+		for dir in list_entries(root, &current_prefix, EntryKind::Directory).await? {
+			to_visit.push(dir);
+		}
+	}
+
+	Ok(paths)
+}
+
+fn is_not_found(error: &JsValue) -> bool {
+	error.dyn_ref::<DomException>().is_some_and(|error| error.name() == "NotFoundError")
+}
+
+/// OPFS raises `TypeMismatchError` when a path segment used as a directory is actually a file.
+fn is_type_mismatch(error: &JsValue) -> bool {
+	error.dyn_ref::<DomException>().is_some_and(|error| error.name() == "TypeMismatchError")
+}
+
+/// Map a listing error: a prefix that names a file becomes [`ContainerError::NotADirectory`], anything
+/// else passes through as a backend error.
+fn list_error(prefix: &str, error: JsValue) -> ContainerError {
+	if is_type_mismatch(&error) {
+		ContainerError::NotADirectory(prefix.to_string())
+	} else {
+		js_err(error)
+	}
+}

document/container/src/lib.rs

@@ -0,0 +1,500 @@
+//! Container abstraction for the on-disk side of the `.gdd` document format.
+//!
+//! A [`Container`] is a virtual filesystem of named byte payloads.
+//! Backends include a loose folder, an in-memory map, and an OPFS-backed wasm store.
+//! Archive codecs ([`archive::Zip`], [`archive::Xz`]) round-trip a container's contents
+//! through a compressed byte stream.
+//!
+//! Reads return a [`ByteHolder`], an ownership-carrying handle whose variant depends on
+//! how the backend produced the bytes (mmap region, owned vector, external file mmap).
+//! [`AsyncContainer`] mirrors [`Container`] for inherently async backends; every sync
+//! [`Container`] is reachable from async code via a blanket impl.
+
+pub mod archive;
+pub mod backends;
+
+pub enum ByteHolder {
+	/// Bytes synthesized in memory (decompressed from an archive, produced by serialization).
+	/// The only variant available on `target_family = "wasm"`.
+	Owned(Vec<u8>),
+	/// Bytes mmap'd from a file inside the container.
+	#[cfg(not(target_family = "wasm"))]
+	Mmapped(MmappedBytes),
+	/// Bytes mmap'd from a file outside the container (e.g. a linked resource).
+	#[cfg(not(target_family = "wasm"))]
+	External { path: std::path::PathBuf, bytes: MmappedBytes },
+}
+
+impl ByteHolder {
+	pub fn as_slice(&self) -> &[u8] {
+		match self {
+			ByteHolder::Owned(bytes) => bytes,
+			#[cfg(not(target_family = "wasm"))]
+			ByteHolder::Mmapped(bytes) => bytes.as_ref(),
+			#[cfg(not(target_family = "wasm"))]
+			ByteHolder::External { bytes, .. } => bytes.as_ref(),
+		}
+	}
+
+	/// If the bytes are backed by a real filesystem path, return it. Enables consumers to
+	/// short-circuit byte copies with `fs::copy` (CoW on supported filesystems).
+	#[cfg(not(target_family = "wasm"))]
+	pub fn source_path(&self) -> Option<&std::path::Path> {
+		match self {
+			ByteHolder::Owned(_) => None,
+			ByteHolder::Mmapped(bytes) => Some(bytes.path()),
+			ByteHolder::External { path, .. } => Some(path),
+		}
+	}
+
+	#[cfg(target_family = "wasm")]
+	pub fn source_path(&self) -> Option<&std::path::Path> {
+		None
+	}
+
+	/// Open an external file and produce a [`ByteHolder::External`] backed by mmap.
+	#[cfg(not(target_family = "wasm"))]
+	pub fn open_external(path: impl Into<std::path::PathBuf>) -> Result<Self> {
+		let path = path.into();
+		let file = mmap_io::mmap::MemoryMappedFile::open_ro(&path).map_err(|error| ContainerError::Backend(format!("mmap of {path:?} failed: {error}")))?;
+		let bytes = MmappedBytes::new(file)?;
+		Ok(ByteHolder::External { path, bytes })
+	}
+}
+
+impl AsRef<[u8]> for ByteHolder {
+	fn as_ref(&self) -> &[u8] {
+		self.as_slice()
+	}
+}
+
+impl std::fmt::Debug for ByteHolder {
+	fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+		match self {
+			ByteHolder::Owned(bytes) => f.debug_tuple("Owned").field(&format_args!("{} bytes", bytes.len())).finish(),
+			#[cfg(not(target_family = "wasm"))]
+			ByteHolder::Mmapped(bytes) => f.debug_tuple("Mmapped").field(&format_args!("{} bytes", bytes.as_ref().len())).finish(),
+			#[cfg(not(target_family = "wasm"))]
+			ByteHolder::External { path, bytes } => f.debug_struct("External").field("path", path).field("len", &bytes.as_ref().len()).finish(),
+		}
+	}
+}
+
+/// Owning wrapper around a memory-mapped file that exposes the mapped region as `&[u8]`.
+#[cfg(not(target_family = "wasm"))]
+pub struct MmappedBytes(mmap_io::mmap::MemoryMappedFile);
+
+#[cfg(not(target_family = "wasm"))]
+impl MmappedBytes {
+	/// Wrap a mapped file, probing that its region is sliceable so the failure surfaces here rather than
+	/// later degrading to the `&[]` fallback in [`AsRef::as_ref`], which cannot return an error.
+	pub fn new(file: mmap_io::mmap::MemoryMappedFile) -> Result<Self> {
+		let len = file.len();
+		file.as_slice(0, len)
+			.map_err(|error| ContainerError::Backend(format!("mmap slice of {:?} failed: {error}", file.path())))?;
+		Ok(Self(file))
+	}
+
+	pub fn path(&self) -> &std::path::Path {
+		self.0.path()
+	}
+}
+
+#[cfg(not(target_family = "wasm"))]
+impl AsRef<[u8]> for MmappedBytes {
+	fn as_ref(&self) -> &[u8] {
+		let len = self.0.len();
+		match self.0.as_slice(0, len) {
+			Ok(slice) => slice,
+			Err(error) => {
+				log::error!("Failed to obtain mmap slice: {error}");
+				&[]
+			}
+		}
+	}
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum ContainerError {
+	#[error("path not found: {0}")]
+	NotFound(String),
+
+	#[error("invalid path: {0}")]
+	InvalidPath(String),
+
+	#[error("not a directory: {0}")]
+	NotADirectory(String),
+
+	#[error("I/O error: {0}")]
+	Io(#[from] std::io::Error),
+
+	/// An archive declared more decompressed data (`declared` bytes) than the codec is willing to
+	/// materialize (`limit` bytes).
+	#[error("declared decompressed size {declared} exceeds the {limit}-byte limit")]
+	SizeLimitExceeded { declared: u64, limit: u64 },
+
+	/// Failure inside an archive codec (zip, lzma, tar). Wraps a foreign error as text since those
+	/// error types don't share a common Rust trait we can chain through.
+	#[error("codec error: {0}")]
+	Codec(String),
+
+	/// Failure inside a storage backend (mmap, OPFS/JS). Wraps a foreign error as text for the same reason.
+	#[error("backend error: {0}")]
+	Backend(String),
+}
+
+pub type Result<T> = std::result::Result<T, ContainerError>;
+
+/// Normalize a listing prefix to end with a trailing slash (unless it names the container root).
+/// The root (empty string or `.`) normalizes to the empty string, so backends concatenate child paths
+/// as `prefix/child` without a double, missing, or `./`-rooted slash.
+pub(crate) fn with_trailing_slash(prefix: &str) -> String {
+	if prefix.is_empty() || prefix == "." {
+		String::new()
+	} else if prefix.ends_with('/') {
+		prefix.to_string()
+	} else {
+		format!("{prefix}/")
+	}
+}
+
+/// Validate that `path` names a container-safe file: relative, no `.`/`..` segments, no backslashes,
+/// no redundant separators (`a//b`, `a/b/`). Dotfile names like `.gitignore` are fine. Requiring a canonical
+/// form gives a file one identity across backends, some of which key on the raw string rather than path
+/// components. Used by backends to block container escapes and by archive codecs on untrusted entry names.
+/// For listing prefixes, which may name the container root, use [`validate_prefix`] instead.
+pub fn validate_path(path: &str) -> Result<()> {
+	let invalid = || ContainerError::InvalidPath(path.to_string());
+
+	if path.is_empty() || path.contains('\\') || path.starts_with('/') {
+		return Err(invalid());
+	}
+
+	// `Path::components` silently folds away `//`, trailing `/`, and interior `.` segments, but backends key
+	// on the raw string, so a non-canonical path would resolve to one file on a path-joining backend yet a
+	// different identity on a string-keyed backend. Reject the redundant segments the component loop below
+	// can't see (it never observes a folded-away `CurDir`/empty segment).
+	if path.split('/').any(|segment| segment.is_empty() || segment == ".") {
+		return Err(invalid());
+	}
+
+	// Reject Windows drive-letter prefixes (`C:foo`, `C:/foo`) that platform-agnostic Path doesn't recognize as absolute on Linux.
+	let bytes = path.as_bytes();
+	if bytes.len() >= 2 && bytes[1] == b':' && bytes[0].is_ascii_alphabetic() {
+		return Err(invalid());
+	}
+
+	for component in std::path::Path::new(path).components() {
+		use std::path::Component;
+		match component {
+			Component::Normal(_) => {}
+			Component::CurDir | Component::ParentDir | Component::Prefix(_) | Component::RootDir => return Err(invalid()),
+		}
+	}
+	Ok(())
+}
+
+/// Validate a listing prefix. Same rules as [`validate_path`], except the container root is also a valid
+/// prefix, named by either the empty string or `.`. Backends pass this to `list`/`list_dirs`.
+pub fn validate_prefix(prefix: &str) -> Result<()> {
+	if prefix.is_empty() || prefix == "." {
+		return Ok(());
+	}
+	validate_path(prefix)
+}
+
+/// Synchronous virtual filesystem of named byte payloads.
+pub trait Container {
+	/// Read the contents of `path` into a [`ByteHolder`].
+	fn read(&self, path: &str) -> Result<ByteHolder>;
+
+	/// Write `bytes` at `path`, creating intermediate directories as needed.
+	fn write(&self, path: &str, bytes: &[u8]) -> Result<()>;
+
+	/// Append `bytes` to the file at `path`, creating it (and any intermediate directories)
+	/// if it does not yet exist. Equivalent to `write` on a fresh path.
+	fn append(&self, path: &str, bytes: &[u8]) -> Result<()>;
+
+	/// Write `size` bytes whose contents are produced by `fill`.
+	/// The default implementation allocates and forwards to [`Container::write`];
+	/// backends that can mmap a writable region may override to fill in place.
+	fn write_sized(&self, path: &str, size: usize, fill: &mut dyn FnMut(&mut [u8]) -> Result<()>) -> Result<()> {
+		let mut buffer = vec![0; size];
+		fill(&mut buffer)?;
+		self.write(path, &buffer)
+	}
+
+	/// List file entries directly under `prefix` (non-recursive).
+	/// Returned paths include the `prefix` (e.g. `list("resources")` returns
+	/// `["resources/abc123", ...]`). A missing prefix yields an empty list; a prefix that names a
+	/// file (not a directory) is an error.
+	fn list(&self, prefix: &str) -> Result<Vec<String>>;
+
+	/// List subdirectory entries directly under `prefix` (non-recursive).
+	/// Returned names include the `prefix`, without a trailing slash
+	/// (e.g. `list_dirs("")` returns `["resources"]`). Same missing/file-prefix semantics as [`Container::list`].
+	fn list_dirs(&self, prefix: &str) -> Result<Vec<String>>;
+
+	/// Whether a file exists at `path`. Directories return `false`.
+	fn exists(&self, path: &str) -> bool;
+
+	/// Remove the file at `path`. Idempotent: removing a missing path succeeds. Directories are never
+	/// removed (they exist only implicitly as parents of files).
+	fn remove(&self, path: &str) -> Result<()>;
+}
+
+/// Asynchronous virtual filesystem of named byte payloads. Mirrors [`Container`].
+///
+/// The returned futures are intentionally not `Send`: native uses `block_on` at the save seam
+/// and wasm is single-threaded, so neither needs cross-thread futures. Revisit if we ever want
+/// to run container I/O on a thread pool.
+#[expect(async_fn_in_trait, reason = "see trait docs — Send is not required")]
+pub trait AsyncContainer {
+	async fn read(&self, path: &str) -> Result<ByteHolder>;
+
+	async fn write(&self, path: &str, bytes: &[u8]) -> Result<()>;
+
+	async fn append(&self, path: &str, bytes: &[u8]) -> Result<()>;
+
+	async fn write_sized(&self, path: &str, size: usize, fill: &mut dyn FnMut(&mut [u8]) -> Result<()>) -> Result<()>;
+
+	async fn list(&self, prefix: &str) -> Result<Vec<String>>;
+
+	async fn list_dirs(&self, prefix: &str) -> Result<Vec<String>>;
+
+	async fn exists(&self, path: &str) -> bool;
+
+	async fn remove(&self, path: &str) -> Result<()>;
+
+	/// Synchronous write. On backends with sync I/O (folder, memory) the write completes durably
+	/// before return and reports real errors. On OPFS the write is enqueued onto a background task
+	/// and `Ok` is returned eagerly; a later failure is logged.
+	fn write_non_blocking(&self, path: &str, bytes: &[u8]) -> Result<()>;
+
+	/// Synchronous append. Same eager-enqueue semantics on OPFS as [`write_non_blocking`](Self::write_non_blocking);
+	/// queued appends preserve order relative to earlier queued writes/appends.
+	fn append_non_blocking(&self, path: &str, bytes: &[u8]) -> Result<()>;
+
+	/// Synchronous remove. Same semantics as [`write_non_blocking`](Self::write_non_blocking).
+	fn remove_non_blocking(&self, path: &str) -> Result<()>;
+
+	/// Non-blocking existence check. On OPFS this reads from an in-memory tracking set populated by
+	/// the sync write/remove paths, since the underlying OPFS existence API is async only.
+	fn exists_non_blocking(&self, path: &str) -> bool;
+}
+
+impl<C: Container + ?Sized> AsyncContainer for C {
+	async fn read(&self, path: &str) -> Result<ByteHolder> {
+		Container::read(self, path)
+	}
+
+	async fn write(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		Container::write(self, path, bytes)
+	}
+
+	async fn append(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		Container::append(self, path, bytes)
+	}
+
+	async fn write_sized(&self, path: &str, size: usize, fill: &mut dyn FnMut(&mut [u8]) -> Result<()>) -> Result<()> {
+		Container::write_sized(self, path, size, fill)
+	}
+
+	async fn list(&self, prefix: &str) -> Result<Vec<String>> {
+		Container::list(self, prefix)
+	}
+
+	async fn list_dirs(&self, prefix: &str) -> Result<Vec<String>> {
+		Container::list_dirs(self, prefix)
+	}
+
+	async fn exists(&self, path: &str) -> bool {
+		Container::exists(self, path)
+	}
+
+	async fn remove(&self, path: &str) -> Result<()> {
+		Container::remove(self, path)
+	}
+
+	fn write_non_blocking(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		Container::write(self, path, bytes)
+	}
+
+	fn append_non_blocking(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		Container::append(self, path, bytes)
+	}
+
+	fn remove_non_blocking(&self, path: &str) -> Result<()> {
+		Container::remove(self, path)
+	}
+
+	fn exists_non_blocking(&self, path: &str) -> bool {
+		Container::exists(self, path)
+	}
+}
+
+/// Type-erased container that dispatches to one of the in-tree backends.
+///
+/// `AsyncContainer::read` returns `impl Future`, so `dyn AsyncContainer` is not object-safe.
+/// `AnyContainer` is the workaround: `Gdd` holds one of these by value, and the `AsyncContainer`
+/// impl forwards to the active variant.
+pub enum AnyContainer {
+	Memory(backends::memory::MemoryBackend),
+	#[cfg(not(target_family = "wasm"))]
+	Folder(backends::folder::FolderBackend),
+	#[cfg(target_family = "wasm")]
+	Opfs(backends::opfs::OpfsBackend),
+}
+
+impl AsyncContainer for AnyContainer {
+	async fn read(&self, path: &str) -> Result<ByteHolder> {
+		match self {
+			Self::Memory(backend) => AsyncContainer::read(backend, path).await,
+			#[cfg(not(target_family = "wasm"))]
+			Self::Folder(backend) => AsyncContainer::read(backend, path).await,
+			#[cfg(target_family = "wasm")]
+			Self::Opfs(backend) => AsyncContainer::read(backend, path).await,
+		}
+	}
+
+	async fn write(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		match self {
+			Self::Memory(backend) => AsyncContainer::write(backend, path, bytes).await,
+			#[cfg(not(target_family = "wasm"))]
+			Self::Folder(backend) => AsyncContainer::write(backend, path, bytes).await,
+			#[cfg(target_family = "wasm")]
+			Self::Opfs(backend) => AsyncContainer::write(backend, path, bytes).await,
+		}
+	}
+
+	async fn append(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		match self {
+			Self::Memory(backend) => AsyncContainer::append(backend, path, bytes).await,
+			#[cfg(not(target_family = "wasm"))]
+			Self::Folder(backend) => AsyncContainer::append(backend, path, bytes).await,
+			#[cfg(target_family = "wasm")]
+			Self::Opfs(backend) => AsyncContainer::append(backend, path, bytes).await,
+		}
+	}
+
+	async fn write_sized(&self, path: &str, size: usize, fill: &mut dyn FnMut(&mut [u8]) -> Result<()>) -> Result<()> {
+		match self {
+			Self::Memory(backend) => AsyncContainer::write_sized(backend, path, size, fill).await,
+			#[cfg(not(target_family = "wasm"))]
+			Self::Folder(backend) => AsyncContainer::write_sized(backend, path, size, fill).await,
+			#[cfg(target_family = "wasm")]
+			Self::Opfs(backend) => AsyncContainer::write_sized(backend, path, size, fill).await,
+		}
+	}
+
+	async fn list(&self, prefix: &str) -> Result<Vec<String>> {
+		match self {
+			Self::Memory(backend) => AsyncContainer::list(backend, prefix).await,
+			#[cfg(not(target_family = "wasm"))]
+			Self::Folder(backend) => AsyncContainer::list(backend, prefix).await,
+			#[cfg(target_family = "wasm")]
+			Self::Opfs(backend) => AsyncContainer::list(backend, prefix).await,
+		}
+	}
+
+	async fn list_dirs(&self, prefix: &str) -> Result<Vec<String>> {
+		match self {
+			Self::Memory(backend) => AsyncContainer::list_dirs(backend, prefix).await,
+			#[cfg(not(target_family = "wasm"))]
+			Self::Folder(backend) => AsyncContainer::list_dirs(backend, prefix).await,
+			#[cfg(target_family = "wasm")]
+			Self::Opfs(backend) => AsyncContainer::list_dirs(backend, prefix).await,
+		}
+	}
+
+	async fn exists(&self, path: &str) -> bool {
+		match self {
+			Self::Memory(backend) => AsyncContainer::exists(backend, path).await,
+			#[cfg(not(target_family = "wasm"))]
+			Self::Folder(backend) => AsyncContainer::exists(backend, path).await,
+			#[cfg(target_family = "wasm")]
+			Self::Opfs(backend) => AsyncContainer::exists(backend, path).await,
+		}
+	}
+
+	async fn remove(&self, path: &str) -> Result<()> {
+		match self {
+			Self::Memory(backend) => AsyncContainer::remove(backend, path).await,
+			#[cfg(not(target_family = "wasm"))]
+			Self::Folder(backend) => AsyncContainer::remove(backend, path).await,
+			#[cfg(target_family = "wasm")]
+			Self::Opfs(backend) => AsyncContainer::remove(backend, path).await,
+		}
+	}
+
+	fn write_non_blocking(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		match self {
+			Self::Memory(backend) => AsyncContainer::write_non_blocking(backend, path, bytes),
+			#[cfg(not(target_family = "wasm"))]
+			Self::Folder(backend) => AsyncContainer::write_non_blocking(backend, path, bytes),
+			#[cfg(target_family = "wasm")]
+			Self::Opfs(backend) => AsyncContainer::write_non_blocking(backend, path, bytes),
+		}
+	}
+
+	fn append_non_blocking(&self, path: &str, bytes: &[u8]) -> Result<()> {
+		match self {
+			Self::Memory(backend) => AsyncContainer::append_non_blocking(backend, path, bytes),
+			#[cfg(not(target_family = "wasm"))]
+			Self::Folder(backend) => AsyncContainer::append_non_blocking(backend, path, bytes),
+			#[cfg(target_family = "wasm")]
+			Self::Opfs(backend) => AsyncContainer::append_non_blocking(backend, path, bytes),
+		}
+	}
+
+	fn remove_non_blocking(&self, path: &str) -> Result<()> {
+		match self {
+			Self::Memory(backend) => AsyncContainer::remove_non_blocking(backend, path),
+			#[cfg(not(target_family = "wasm"))]
+			Self::Folder(backend) => AsyncContainer::remove_non_blocking(backend, path),
+			#[cfg(target_family = "wasm")]
+			Self::Opfs(backend) => AsyncContainer::remove_non_blocking(backend, path),
+		}
+	}
+
+	fn exists_non_blocking(&self, path: &str) -> bool {
+		match self {
+			Self::Memory(backend) => AsyncContainer::exists_non_blocking(backend, path),
+			#[cfg(not(target_family = "wasm"))]
+			Self::Folder(backend) => AsyncContainer::exists_non_blocking(backend, path),
+			#[cfg(target_family = "wasm")]
+			Self::Opfs(backend) => AsyncContainer::exists_non_blocking(backend, path),
+		}
+	}
+}
+
+#[cfg(test)]
+mod tests {
+	use super::{ContainerError, validate_path, validate_prefix};
+
+	#[test]
+	fn validate_path_accepts_well_formed() {
+		// `..` and `.` are only rejected as whole path components; as substrings of a name they are fine.
+		for ok in ["manifest.json", "resources/abc", "a/b/c.bin", "my..file.txt", "..hidden", ".gitignore"] {
+			assert!(validate_path(ok).is_ok(), "{ok:?} should be accepted");
+		}
+	}
+
+	#[test]
+	fn validate_path_rejects_unsafe() {
+		for bad in ["", "../escape", "a/../b", ".", "./leading", "/abs", "back\\slash", "C:/win", "\\\\?\\unc"] {
+			let result = validate_path(bad);
+			assert!(matches!(result, Err(ContainerError::InvalidPath(_))), "{bad:?} should be rejected, got {result:?}");
+		}
+	}
+
+	#[test]
+	fn validate_prefix_accepts_root_tokens() {
+		// The container root is a valid listing prefix, named by either the empty string or `.`.
+		for root in ["", ".", "resources", "a/b"] {
+			assert!(validate_prefix(root).is_ok(), "{root:?} should be accepted as a prefix");
+		}
+		// Unsafe prefixes are still rejected, same as paths.
+		assert!(validate_prefix("../escape").is_err());
+	}
+}

document/container/tests/archive.rs

@@ -0,0 +1,81 @@
+#![cfg(any(feature = "zip", feature = "xz"))]
+
+use document_container::Container;
+use document_container::archive::{Archive, ArchiveWriter};
+use document_container::backends::folder::FolderBackend;
+use document_container::backends::memory::MemoryBackend;
+
+fn entries() -> Vec<(&'static str, &'static [u8])> {
+	vec![
+		("manifest.json", br#"{"format":"gdd"}"#),
+		("document.json", b"{\"registry\":\"...\"}"),
+		("history.jsonl", b"{\"rev\":1}\n{\"rev\":2}\n"),
+		("resources/abc123", &[0xDE, 0xAD, 0xBE, 0xEF]),
+		("resources/xyz789", b"another resource"),
+	]
+}
+
+fn assert_round_trip(restored: &MemoryBackend) {
+	assert_eq!(restored.read("manifest.json").unwrap().as_slice(), br#"{"format":"gdd"}"#);
+	assert_eq!(restored.read("document.json").unwrap().as_slice(), b"{\"registry\":\"...\"}");
+	assert_eq!(restored.read("history.jsonl").unwrap().as_slice(), b"{\"rev\":1}\n{\"rev\":2}\n");
+	assert_eq!(restored.read("resources/abc123").unwrap().as_slice(), &[0xDE, 0xAD, 0xBE, 0xEF]);
+	assert_eq!(restored.read("resources/xyz789").unwrap().as_slice(), b"another resource");
+}
+
+#[cfg(feature = "zip")]
+#[test]
+fn zip_round_trip() {
+	use document_container::archive::Zip;
+	use std::io::Cursor;
+
+	let mut buffer = Cursor::new(Vec::new());
+	let mut writer = Zip::writer(&mut buffer).unwrap();
+	for (path, bytes) in entries() {
+		writer.write_entry(path, bytes).unwrap();
+	}
+	writer.finish().unwrap();
+
+	let mut restored = MemoryBackend::new();
+	<Zip as Archive>::open(Cursor::new(buffer.get_ref()), &mut restored).unwrap();
+	assert_round_trip(&restored);
+}
+
+#[cfg(feature = "zip")]
+#[test]
+fn zip_deserialize_streams_into_folder_backend() {
+	use document_container::archive::Zip;
+	use std::io::Cursor;
+
+	let mut buffer = Cursor::new(Vec::new());
+	let mut writer = Zip::writer(&mut buffer).unwrap();
+	for (path, bytes) in entries() {
+		writer.write_entry(path, bytes).unwrap();
+	}
+	writer.finish().unwrap();
+
+	let dir = tempfile::tempdir().unwrap();
+	let mut restored = FolderBackend::create(dir.path()).unwrap();
+	<Zip as Archive>::open(Cursor::new(buffer.get_ref()), &mut restored).unwrap();
+
+	assert_eq!(restored.read("manifest.json").unwrap().as_slice(), br#"{"format":"gdd"}"#);
+	assert_eq!(restored.read("resources/abc123").unwrap().as_slice(), &[0xDE, 0xAD, 0xBE, 0xEF]);
+}
+
+#[cfg(feature = "xz")]
+#[test]
+fn xz_round_trip() {
+	use document_container::archive::Xz;
+	use std::io::Cursor;
+
+	let mut buffer = Cursor::new(Vec::new());
+	let mut writer = Xz::writer(&mut buffer).unwrap();
+	for (path, bytes) in entries() {
+		writer.write_entry(path, bytes).unwrap();
+	}
+	writer.finish().unwrap();
+
+	let mut restored = MemoryBackend::new();
+	<Xz as Archive>::open(Cursor::new(buffer.get_ref()), &mut restored).unwrap();
+	assert_round_trip(&restored);
+}

document/container/tests/round_trip.rs

@@ -0,0 +1,178 @@
+use document_container::backends::folder::FolderBackend;
+use document_container::backends::memory::MemoryBackend;
+use document_container::{AnyContainer, Container, ContainerError};
+
+fn run_round_trip<C: Container>(container: C) {
+	container.write("manifest.json", br#"{"format":"gdd"}"#).unwrap();
+	container.write("resources/abc123", &[0xDE, 0xAD, 0xBE, 0xEF]).unwrap();
+	container.write("resources/xyz789", b"another resource").unwrap();
+
+	assert!(container.exists("manifest.json"));
+	assert!(container.exists("resources/abc123"));
+	assert!(!container.exists("does-not-exist"));
+
+	let manifest = container.read("manifest.json").unwrap();
+	assert_eq!(manifest.as_slice(), br#"{"format":"gdd"}"#);
+
+	let blob = container.read("resources/abc123").unwrap();
+	assert_eq!(blob.as_slice(), &[0xDE, 0xAD, 0xBE, 0xEF]);
+
+	let top_level = container.list("").unwrap();
+	assert!(top_level.iter().any(|p| p == "manifest.json"));
+	assert!(
+		top_level.iter().all(|p| !p.starts_with("resources/")),
+		"list(\"\") must not descend into subdirectories, got {top_level:?}"
+	);
+
+	let mut resources = container.list("resources").unwrap();
+	resources.sort();
+	assert_eq!(resources, vec!["resources/abc123".to_string(), "resources/xyz789".to_string()]);
+
+	container.remove("resources/abc123").unwrap();
+	assert!(!container.exists("resources/abc123"));
+	assert!(matches!(container.read("resources/abc123"), Err(ContainerError::NotFound(_))));
+}
+
+#[test]
+fn memory_backend_round_trip() {
+	run_round_trip(MemoryBackend::new());
+}
+
+#[test]
+fn folder_backend_round_trip() {
+	let dir = tempfile::tempdir().unwrap();
+	let backend = FolderBackend::create(dir.path()).unwrap();
+	run_round_trip(backend);
+}
+
+fn run_list_and_remove_semantics<C: Container>(container: C) {
+	container.write("dir/file", b"x").unwrap();
+
+	// Removing a missing path is idempotent.
+	container.remove("dir/missing").unwrap();
+	container.remove("dir/file").unwrap();
+	container.remove("dir/file").unwrap();
+
+	// Listing a missing prefix yields an empty list.
+	assert_eq!(container.list("nonexistent").unwrap(), Vec::<String>::new());
+
+	// Listing a prefix that names a file is an error.
+	container.write("manifest.json", b"{}").unwrap();
+	assert!(matches!(container.list("manifest.json"), Err(ContainerError::NotADirectory(_))));
+	assert!(matches!(container.list_dirs("manifest.json"), Err(ContainerError::NotADirectory(_))));
+}
+
+#[test]
+fn memory_backend_list_and_remove_semantics() {
+	run_list_and_remove_semantics(MemoryBackend::new());
+}
+
+#[test]
+fn folder_backend_list_and_remove_semantics() {
+	let dir = tempfile::tempdir().unwrap();
+	run_list_and_remove_semantics(FolderBackend::create(dir.path()).unwrap());
+}
+
+#[test]
+#[cfg(unix)]
+fn folder_backend_rejects_symlink_escape() {
+	let outside = tempfile::tempdir().unwrap();
+	std::fs::write(outside.path().join("secret"), b"sensitive").unwrap();
+
+	let dir = tempfile::tempdir().unwrap();
+	let backend = FolderBackend::create(dir.path()).unwrap();
+	// A symlink planted inside the root pointing outside it must not be traversable.
+	std::os::unix::fs::symlink(outside.path(), dir.path().join("link")).unwrap();
+
+	let result = backend.read("link/secret");
+	assert!(matches!(result, Err(ContainerError::InvalidPath(_))), "symlink escape should be rejected, got {result:?}");
+
+	// Listing must reject the symlinked prefix too, not just reads, so it can't leak entry names from outside the root.
+	assert!(matches!(backend.list("link"), Err(ContainerError::InvalidPath(_))), "list through symlink should be rejected");
+	assert!(matches!(backend.list_dirs("link"), Err(ContainerError::InvalidPath(_))), "list_dirs through symlink should be rejected");
+
+	// A symlink entry directly under root must not appear in listings, since `resolve` would reject reading it.
+	std::os::unix::fs::symlink(outside.path().join("secret"), dir.path().join("file_link")).unwrap();
+	std::fs::write(dir.path().join("real"), b"ok").unwrap();
+	assert_eq!(backend.list("").unwrap(), vec!["real".to_string()], "symlink entries should be omitted from file listings");
+	assert!(backend.list_dirs("").unwrap().is_empty(), "symlink-to-dir entries should be omitted from dir listings");
+}
+
+#[test]
+fn folder_backend_rejects_path_traversal() {
+	let dir = tempfile::tempdir().unwrap();
+	let backend = FolderBackend::create(dir.path()).unwrap();
+	for bad in ["../escape", "subdir/../escape", "/abs", "back\\slash", "a//b", "trailing/", "a/./b"] {
+		let result = backend.write(bad, b"nope");
+		assert!(matches!(result, Err(ContainerError::InvalidPath(_))), "expected InvalidPath for {bad:?}, got {result:?}");
+	}
+}
+
+fn run_append<C: Container>(container: C) {
+	// Appending to a non-existent path creates it — same semantics as `OpenOptions::append().create(true)`.
+	container.append("history.jsonl", b"{\"op\":1}\n").unwrap();
+	container.append("history.jsonl", b"{\"op\":2}\n").unwrap();
+	container.append("history.jsonl", b"{\"op\":3}\n").unwrap();
+
+	let log = container.read("history.jsonl").unwrap();
+	assert_eq!(log.as_slice(), b"{\"op\":1}\n{\"op\":2}\n{\"op\":3}\n");
+}
+
+#[test]
+fn memory_backend_append() {
+	run_append(MemoryBackend::new());
+}
+
+#[test]
+fn folder_backend_append() {
+	let dir = tempfile::tempdir().unwrap();
+	let backend = FolderBackend::create(dir.path()).unwrap();
+	run_append(backend);
+}
+
+#[test]
+fn any_container_dispatches_to_active_variant() {
+	use document_container::AsyncContainer;
+
+	let container = AnyContainer::Memory(MemoryBackend::new());
+
+	futures::executor::block_on(async {
+		container.write("manifest.json", br#"{"format":"gdd"}"#).await.unwrap();
+		container.append("history.jsonl", b"frame-1\n").await.unwrap();
+		container.append("history.jsonl", b"frame-2\n").await.unwrap();
+
+		assert!(container.exists("manifest.json").await);
+		let manifest = container.read("manifest.json").await.unwrap();
+		assert_eq!(manifest.as_slice(), br#"{"format":"gdd"}"#);
+
+		let history = container.read("history.jsonl").await.unwrap();
+		assert_eq!(history.as_slice(), b"frame-1\nframe-2\n");
+	});
+}
+
+#[test]
+fn folder_backend_reads_empty_file() {
+	let dir = tempfile::tempdir().unwrap();
+	let backend = FolderBackend::create(dir.path()).unwrap();
+
+	backend.write("empty.bin", &[]).unwrap();
+	let read_back = backend.read("empty.bin").unwrap();
+	assert_eq!(read_back.as_slice(), &[] as &[u8]);
+}
+
+#[test]
+fn folder_backend_write_sized_fills_via_mmap() {
+	let dir = tempfile::tempdir().unwrap();
+	let backend = FolderBackend::create(dir.path()).unwrap();
+
+	let payload = b"hello world";
+	backend
+		.write_sized("resources/sized", payload.len(), &mut |buffer| {
+			buffer.copy_from_slice(payload);
+			Ok(())
+		})
+		.unwrap();
+
+	let read_back = backend.read("resources/sized").unwrap();
+	assert_eq!(read_back.as_slice(), payload);
+}