restore.py

#!/usr/bin/env python3
"""Safely verify and restore independently compressed RoboDojo Assets shards."""

from __future__ import annotations

import argparse
import concurrent.futures
import ctypes
import errno
import fcntl
import hashlib
import json
import os
from pathlib import Path, PurePosixPath
import shutil
import stat
import subprocess
import sys
import tarfile
from typing import Any, Dict, Iterable, List, Optional, Sequence, Tuple
import uuid


READ_BLOCK = 8 * 1024 * 1024
RENAME_NOREPLACE = 1
RENAME_EXCHANGE = 2


def renameat2(src_dir_fd: int, src: str, dst_dir_fd: int, dst: str, flags: int) -> None:
    libc = ctypes.CDLL(None, use_errno=True)
    function = getattr(libc, "renameat2", None)
    if function is None:
        raise OSError(errno.ENOSYS, "libc does not expose renameat2", dst)
    function.argtypes = [ctypes.c_int, ctypes.c_char_p, ctypes.c_int, ctypes.c_char_p, ctypes.c_uint]
    function.restype = ctypes.c_int
    result = function(src_dir_fd, os.fsencode(src), dst_dir_fd, os.fsencode(dst), flags)
    if result != 0:
        error_number = ctypes.get_errno()
        raise OSError(error_number, os.strerror(error_number), dst)


class PublicationRecoveryError(RuntimeError):
    """The old tree may remain in the private staging directory."""


class PublicationRolledBack(RuntimeError):
    """Publication failed, but the original tree was restored safely."""

    def __init__(self, original: BaseException):
        super().__init__(f"publication rolled back after: {original}")
        self.original = original


def renameat2_unsupported(error: OSError) -> bool:
    return error.errno in {
        errno.EINVAL,
        errno.ENOSYS,
        getattr(errno, "EOPNOTSUPP", errno.EINVAL),
        getattr(errno, "ENOTSUP", errno.EINVAL),
    }


def publish_new(final_stage_fd: int, output_fd: int) -> None:
    try:
        renameat2(final_stage_fd, "Assets", output_fd, "Assets", RENAME_NOREPLACE)
    except OSError as error:
        if not renameat2_unsupported(error):
            raise
        if lstat_at(output_fd, "Assets") is not None:
            raise RuntimeError("Assets appeared before fallback publication")
        os.mkdir("Assets", 0o700, dir_fd=output_fd)
        reservation = lstat_at(output_fd, "Assets")
        if reservation is None or not stat.S_ISDIR(reservation.st_mode):
            raise RuntimeError("failed to reserve Assets for fallback publication")
        try:
            current = lstat_at(output_fd, "Assets")
            if current is None or (current.st_dev, current.st_ino) != (reservation.st_dev, reservation.st_ino):
                raise RuntimeError("Assets reservation changed before fallback publication")
            os.rename("Assets", "Assets", src_dir_fd=final_stage_fd, dst_dir_fd=output_fd)
        except BaseException:
            current = lstat_at(output_fd, "Assets")
            if current is not None and (current.st_dev, current.st_ino) == (
                reservation.st_dev,
                reservation.st_ino,
            ):
                try:
                    os.rmdir("Assets", dir_fd=output_fd)
                except OSError:
                    pass
            raise


def publish_force(final_stage_fd: int, output_fd: int, backup_name: str) -> None:
    renameat2(final_stage_fd, "Assets", output_fd, "Assets", RENAME_EXCHANGE)
    try:
        os.rename("Assets", backup_name, src_dir_fd=final_stage_fd, dst_dir_fd=output_fd)
    except BaseException as publication_error:
        try:
            renameat2(final_stage_fd, "Assets", output_fd, "Assets", RENAME_EXCHANGE)
        except BaseException as rollback_error:
            raise PublicationRecoveryError(
                f"atomic publication failed and rollback also failed: {rollback_error}"
            ) from publication_error
        raise PublicationRolledBack(publication_error) from publication_error


def lexists(path: Path) -> bool:
    return os.path.lexists(str(path))


def sha256_fd(fd: int) -> str:
    os.lseek(fd, 0, os.SEEK_SET)
    info = os.fstat(fd)
    if not stat.S_ISREG(info.st_mode):
        raise RuntimeError("file descriptor is not a regular file")
    digest = hashlib.sha256()
    while True:
        block = os.read(fd, READ_BLOCK)
        if not block:
            break
        digest.update(block)
    return digest.hexdigest()


def sha256_regular_file(path: Path) -> str:
    nofollow = getattr(os, "O_NOFOLLOW", 0)
    if not nofollow:
        raise RuntimeError("this restore tool requires O_NOFOLLOW support")
    fd = os.open(str(path), os.O_RDONLY | nofollow)
    try:
        return sha256_fd(fd)
    finally:
        os.close(fd)


def validate_payload_path(name: str) -> Tuple[str, ...]:
    if not name or "\x00" in name or any(ord(char) < 32 or ord(char) == 127 for char in name):
        raise RuntimeError(f"unsafe control character in payload path: {name!r}")
    pure = PurePosixPath(name)
    if pure.is_absolute() or pure.as_posix() != name:
        raise RuntimeError(f"non-canonical payload path: {name!r}")
    parts = pure.parts
    if len(parts) < 2 or parts[0] != "Assets" or any(part in ("", ".", "..") for part in parts):
        raise RuntimeError(f"unsafe payload path: {name!r}")
    return parts


def validate_repo_relative_path(name: str, expected_root: str) -> Tuple[str, ...]:
    if not name or "\x00" in name or any(ord(char) < 32 or ord(char) == 127 for char in name):
        raise RuntimeError(f"unsafe repository path: {name!r}")
    pure = PurePosixPath(name)
    if pure.is_absolute() or pure.as_posix() != name:
        raise RuntimeError(f"non-canonical repository path: {name!r}")
    parts = pure.parts
    if len(parts) != 2 or parts[0] != expected_root or any(part in ("", ".", "..") for part in parts):
        raise RuntimeError(f"invalid {expected_root} path: {name!r}")
    return parts


def require_regular_repo_file(repo_dir: Path, relative: str, expected_root: str) -> Path:
    parts = validate_repo_relative_path(relative, expected_root)
    current = repo_dir
    for part in parts[:-1]:
        current = current / part
        info = os.lstat(str(current))
        if not stat.S_ISDIR(info.st_mode):
            raise RuntimeError(f"repository parent is not a real directory: {current}")
    path = repo_dir.joinpath(*parts)
    info = os.lstat(str(path))
    if not stat.S_ISREG(info.st_mode):
        raise RuntimeError(f"repository entry is not a regular file: {path}")
    return path


def read_regular_at(directory_fd: int, name: str, max_bytes: int = 64 * 1024 * 1024) -> bytes:
    fd = os.open(
        name,
        os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_NONBLOCK", 0),
        dir_fd=directory_fd,
    )
    try:
        info = os.fstat(fd)
        if not stat.S_ISREG(info.st_mode):
            raise RuntimeError(f"repository entry is not a regular file: {name}")
        if info.st_size > max_bytes:
            raise RuntimeError(f"repository metadata is unexpectedly large: {name}")
        chunks: List[bytes] = []
        remaining = max_bytes + 1
        while remaining > 0:
            block = os.read(fd, min(1024 * 1024, remaining))
            if not block:
                break
            chunks.append(block)
            remaining -= len(block)
        raw = b"".join(chunks)
        if len(raw) > max_bytes:
            raise RuntimeError(f"repository metadata exceeds limit: {name}")
        return raw
    finally:
        os.close(fd)


def parse_list0(raw: bytes, label: str) -> List[str]:
    if not raw or not raw.endswith(b"\x00"):
        raise RuntimeError(f"NUL-delimited list must end in NUL: {label}")
    chunks = raw[:-1].split(b"\x00")
    if any(not chunk for chunk in chunks):
        raise RuntimeError(f"empty member in list: {label}")
    names = [chunk.decode("utf-8", errors="strict") for chunk in chunks]
    for name in names:
        validate_payload_path(name)
    return names


def parse_file_hashes(raw: bytes, label: str) -> Dict[str, str]:
    entries: Dict[str, str] = {}
    text = raw.decode("utf-8", errors="strict")
    for number, line in enumerate(text.splitlines(), 1):
        if len(line) < 67 or line[64:66] != "  ":
            raise RuntimeError(f"invalid FILES.sha256 line {number}")
        digest = line[:64]
        name = line[66:]
        if digest != digest.lower() or any(char not in "0123456789abcdef" for char in digest):
            raise RuntimeError(f"invalid SHA-256 digest on line {number}")
        validate_payload_path(name)
        if name in entries:
            raise RuntimeError(f"duplicate path in FILES.sha256: {name}")
        entries[name] = digest
    if not entries:
        raise RuntimeError("FILES.sha256 is empty")
    return entries


def check_leaf_ancestor_conflicts(names: Iterable[str]) -> None:
    name_set = set(names)
    for name in name_set:
        parts = PurePosixPath(name).parts
        for depth in range(1, len(parts)):
            ancestor = "/".join(parts[:depth])
            if ancestor in name_set:
                raise RuntimeError(f"payload leaf/ancestor conflict: {ancestor!r} and {name!r}")


def load_metadata(
    repo_dir: Path, repo_fd: int
) -> Tuple[Dict[str, Any], List[Dict[str, Any]], Dict[str, str], Dict[str, int]]:
    manifest_raw = read_regular_at(repo_fd, "MANIFEST.json")
    hashes_raw = read_regular_at(repo_fd, "FILES.sha256")
    manifest = json.loads(manifest_raw.decode("utf-8", errors="strict"))
    if manifest.get("format_version") != 1:
        raise RuntimeError("unsupported manifest format_version")
    shards = manifest.get("shards")
    if not isinstance(shards, list) or not shards:
        raise RuntimeError("manifest shards must be a non-empty list")
    if manifest.get("shard_count") != len(shards):
        raise RuntimeError("manifest shard_count mismatch")

    file_hashes = parse_file_hashes(hashes_raw, "FILES.sha256")
    directory_flags = os.O_RDONLY | getattr(os, "O_DIRECTORY", 0) | getattr(os, "O_NOFOLLOW", 0)
    data_fd = os.open("data", directory_flags, dir_fd=repo_fd)
    lists_fd = os.open("lists", directory_flags, dir_fd=repo_fd)
    all_names: List[str] = []
    path_to_shard: Dict[str, int] = {}
    archive_paths = set()
    list_paths = set()
    normalized_shards: List[Dict[str, Any]] = []
    for position, raw_shard in enumerate(shards):
        if not isinstance(raw_shard, dict) or raw_shard.get("index") != position:
            raise RuntimeError(f"invalid shard index at position {position}")
        archive_rel = raw_shard.get("archive")
        list_rel = raw_shard.get("list")
        if not isinstance(archive_rel, str) or not isinstance(list_rel, str):
            raise RuntimeError(f"missing archive/list path for shard {position}")
        if archive_rel in archive_paths or list_rel in list_paths:
            raise RuntimeError("duplicate archive or list path in manifest")
        archive_paths.add(archive_rel)
        list_paths.add(list_rel)
        archive_parts = validate_repo_relative_path(archive_rel, "data")
        list_parts = validate_repo_relative_path(list_rel, "lists")
        archive = repo_dir / archive_rel
        list_path = repo_dir / list_rel
        archive_fd = os.open(
            archive_parts[1],
            os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_NONBLOCK", 0),
            dir_fd=data_fd,
        )
        archive_info = os.fstat(archive_fd)
        if not stat.S_ISREG(archive_info.st_mode):
            os.close(archive_fd)
            raise RuntimeError(f"archive is not a regular file: {archive}")
        names = parse_list0(read_regular_at(lists_fd, list_parts[1]), list_rel)
        if raw_shard.get("file_count") != len(names):
            raise RuntimeError(f"file_count mismatch for shard {position}")
        if not isinstance(raw_shard.get("input_bytes"), int) or raw_shard["input_bytes"] < 0:
            raise RuntimeError(f"invalid input_bytes for shard {position}")
        if not isinstance(raw_shard.get("archive_bytes"), int) or raw_shard["archive_bytes"] <= 0:
            raise RuntimeError(f"invalid archive_bytes for shard {position}")
        digest = raw_shard.get("sha256")
        if not isinstance(digest, str) or len(digest) != 64 or any(
            char not in "0123456789abcdef" for char in digest
        ):
            raise RuntimeError(f"invalid archive SHA-256 for shard {position}")
        for name in names:
            if name in path_to_shard:
                raise RuntimeError(f"payload path appears in multiple shards: {name}")
            path_to_shard[name] = position
        all_names.extend(names)
        normalized = dict(raw_shard)
        normalized["archive_path"] = archive
        normalized["archive_fd"] = archive_fd
        normalized["list_path"] = list_path
        normalized["names"] = names
        normalized_shards.append(normalized)

    os.close(data_fd)
    os.close(lists_fd)

    if set(all_names) != set(file_hashes):
        missing = sorted(set(file_hashes) - set(all_names))[:10]
        extra = sorted(set(all_names) - set(file_hashes))[:10]
        raise RuntimeError(f"list0/FILES.sha256 path mismatch; missing={missing}, extra={extra}")
    if len(all_names) != len(file_hashes):
        raise RuntimeError("payload paths are not globally unique")
    check_leaf_ancestor_conflicts(all_names)
    if manifest.get("source_file_count") != len(file_hashes):
        raise RuntimeError("manifest source_file_count mismatch")
    if sum(int(shard["input_bytes"]) for shard in normalized_shards) != manifest.get("source_bytes"):
        raise RuntimeError("manifest source_bytes mismatch")
    return manifest, normalized_shards, file_hashes, path_to_shard


def validate_tar_member(member: tarfile.TarInfo, expected_name: str) -> None:
    if member.name != expected_name:
        raise RuntimeError(f"archive member mismatch: expected {expected_name!r}, got {member.name!r}")
    validate_payload_path(member.name)
    sparse = getattr(member, "sparse", None)
    sparse_pax = any(str(key).startswith("GNU.sparse.") for key in member.pax_headers)
    if not member.isreg() or member.linkname not in ("", None) or sparse or sparse_pax:
        raise RuntimeError(f"non-regular/link/sparse archive member rejected: {member.name!r}")
    if member.size < 0:
        raise RuntimeError(f"negative archive member size: {member.name!r}")


def open_tar_stream(archive_fd: int) -> Tuple[subprocess.Popen[bytes], tarfile.TarFile]:
    os.lseek(archive_fd, 0, os.SEEK_SET)
    process = subprocess.Popen(
        ["zstd", "--decompress", "--stdout", "--quiet"],
        stdin=archive_fd,
        stdout=subprocess.PIPE,
        stderr=subprocess.PIPE,
    )
    if process.stdout is None:
        process.kill()
        raise RuntimeError("failed to open zstd stdout")
    try:
        stream = tarfile.open(fileobj=process.stdout, mode="r|")
    except Exception:
        process.kill()
        process.communicate()
        raise
    return process, stream


def finish_tar_stream(process: subprocess.Popen[bytes], stream: tarfile.TarFile) -> None:
    stream.close()
    if process.stdout is not None:
        process.stdout.close()
    stderr = process.stderr.read() if process.stderr is not None else b""
    returncode = process.wait()
    if returncode != 0:
        raise RuntimeError(f"zstd failed with code {returncode}: {stderr.decode(errors='replace')[-2000:]}")


def abort_tar_stream(process: subprocess.Popen[bytes], stream: tarfile.TarFile) -> None:
    try:
        stream.close()
    finally:
        process.kill()
        process.communicate()


def preflight_shard(shard: Dict[str, Any]) -> Dict[str, int]:
    archive = Path(shard["archive_path"])
    archive_fd = int(shard["archive_fd"])
    info = os.fstat(archive_fd)
    if not stat.S_ISREG(info.st_mode) or info.st_size != int(shard["archive_bytes"]):
        raise RuntimeError(f"archive size/type mismatch: {archive}")
    if sha256_fd(archive_fd) != shard["sha256"]:
        raise RuntimeError(f"archive SHA-256 mismatch: {archive}")

    process, stream = open_tar_stream(archive_fd)
    expected = list(shard["names"])
    sizes: Dict[str, int] = {}
    try:
        count = 0
        total = 0
        for member in stream:
            if count >= len(expected):
                raise RuntimeError(f"unexpected extra archive member: {member.name!r}")
            validate_tar_member(member, expected[count])
            sizes[member.name] = member.size
            count += 1
            total += member.size
        if count != len(expected):
            raise RuntimeError(f"archive member count mismatch: {archive}")
        if total != int(shard["input_bytes"]):
            raise RuntimeError(f"archive input byte total mismatch: {archive}")
        finish_tar_stream(process, stream)
    except Exception:
        if process.poll() is None:
            abort_tar_stream(process, stream)
        raise
    return sizes


def open_or_create_directory_chain(path: Path) -> int:
    absolute = Path(os.path.abspath(str(path)))
    if absolute.anchor != "/":
        raise RuntimeError(f"output path must be absolute on Linux: {absolute}")
    flags = os.O_RDONLY | getattr(os, "O_DIRECTORY", 0) | getattr(os, "O_NOFOLLOW", 0)
    current_fd = os.open("/", flags)
    for part in absolute.parts[1:]:
        try:
            os.mkdir(part, 0o755, dir_fd=current_fd)
        except FileExistsError:
            pass
        try:
            next_fd = os.open(part, flags, dir_fd=current_fd)
        except Exception:
            os.close(current_fd)
            raise RuntimeError(f"output path component is not a real directory: {absolute}")
        os.close(current_fd)
        current_fd = next_fd
    return current_fd


def open_existing_directory_chain(path: Path) -> int:
    absolute = Path(os.path.abspath(str(path)))
    if absolute.anchor != "/":
        raise RuntimeError(f"repository path must be absolute on Linux: {absolute}")
    flags = os.O_RDONLY | getattr(os, "O_DIRECTORY", 0) | getattr(os, "O_NOFOLLOW", 0)
    current_fd = os.open("/", flags)
    for part in absolute.parts[1:]:
        try:
            next_fd = os.open(part, flags, dir_fd=current_fd)
        except Exception:
            os.close(current_fd)
            raise RuntimeError(f"repository path component is not a real directory: {absolute}")
        os.close(current_fd)
        current_fd = next_fd
    return current_fd


def lstat_at(directory_fd: int, name: str) -> Optional[os.stat_result]:
    try:
        return os.stat(name, dir_fd=directory_fd, follow_symlinks=False)
    except FileNotFoundError:
        return None


def ensure_safe_parents(root: Path, parts: Sequence[str]) -> Path:
    current = root
    for part in parts:
        current = current / part
        try:
            os.mkdir(str(current), 0o755)
        except FileExistsError:
            info = os.lstat(str(current))
            if not stat.S_ISDIR(info.st_mode):
                raise RuntimeError(f"restore parent is not a real directory: {current}")
    return current


def write_member(source: Any, destination: Path, expected_size: int) -> None:
    flags = os.O_WRONLY | os.O_CREAT | os.O_EXCL | getattr(os, "O_NOFOLLOW", 0)
    fd = os.open(str(destination), flags, 0o644)
    written = 0
    try:
        with os.fdopen(fd, "wb") as output:
            fd = -1
            while True:
                block = source.read(READ_BLOCK)
                if not block:
                    break
                output.write(block)
                written += len(block)
    except Exception:
        if fd >= 0:
            os.close(fd)
        try:
            destination.unlink()
        except FileNotFoundError:
            pass
        raise
    if written != expected_size:
        destination.unlink()
        raise RuntimeError(f"short archive member: {destination} ({written} != {expected_size})")


def extract_shard(shard: Dict[str, Any], shard_root: Path) -> int:
    os.mkdir(str(shard_root), 0o700)
    process, stream = open_tar_stream(int(shard["archive_fd"]))
    expected = list(shard["names"])
    total = 0
    try:
        count = 0
        for member in stream:
            if count >= len(expected):
                raise RuntimeError(f"unexpected extra archive member: {member.name!r}")
            validate_tar_member(member, expected[count])
            parts = validate_payload_path(member.name)
            parent = ensure_safe_parents(shard_root, parts[:-1])
            destination = parent / parts[-1]
            source = stream.extractfile(member)
            if source is None:
                raise RuntimeError(f"unable to read regular member: {member.name}")
            with source:
                write_member(source, destination, member.size)
            total += member.size
            count += 1
        if count != len(expected) or total != int(shard["input_bytes"]):
            raise RuntimeError(f"archive changed between preflight and extraction: {shard['archive_path']}")
        finish_tar_stream(process, stream)
    except Exception:
        if process.poll() is None:
            abort_tar_stream(process, stream)
        raise
    return count


def scan_tree(root: Path) -> List[str]:
    names: List[str] = []
    for directory, dirnames, filenames in os.walk(str(root), topdown=True, followlinks=False):
        directory_path = Path(directory)
        for name in dirnames:
            path = directory_path / name
            info = os.lstat(str(path))
            if not stat.S_ISDIR(info.st_mode):
                raise RuntimeError(f"non-directory node in restored tree: {path}")
        for name in filenames:
            path = directory_path / name
            info = os.lstat(str(path))
            if not stat.S_ISREG(info.st_mode):
                raise RuntimeError(f"non-regular node in restored tree: {path}")
            names.append(path.relative_to(root.parent).as_posix())
    return sorted(names)


def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument("--repo-dir", type=Path, default=Path(__file__).absolute().parent)
    parser.add_argument("--output-dir", type=Path, required=True)
    parser.add_argument("--workers", type=int, default=4, help="parallel extraction workers")
    parser.add_argument("--hash-workers", type=int, default=16)
    parser.add_argument("--force", action="store_true", help="atomically replace an existing real Assets directory")
    parser.add_argument("--skip-file-check", action="store_true", help="skip content hashes, never structural checks")
    args = parser.parse_args()

    if shutil.which("zstd") is None:
        parser.error("zstd is required")
    if not getattr(os, "O_NOFOLLOW", 0):
        parser.error("this platform does not provide O_NOFOLLOW")
    if args.workers < 1 or args.hash_workers < 1:
        parser.error("worker counts must be positive")

    repo_dir = Path(os.path.abspath(str(args.repo_dir)))
    repo_fd = open_existing_directory_chain(repo_dir)
    try:
        manifest, shards, file_hashes, path_to_shard = load_metadata(repo_dir, repo_fd)
    finally:
        os.close(repo_fd)
    print(f"Preflighting {len(shards)} archives before creating output...", flush=True)
    all_sizes: Dict[str, int] = {}
    with concurrent.futures.ThreadPoolExecutor(max_workers=min(args.workers, len(shards))) as pool:
        for shard, sizes in zip(shards, pool.map(preflight_shard, shards)):
            all_sizes.update(sizes)
            print(f"  OK {Path(shard['archive_path']).name}", flush=True)
    if set(all_sizes) != set(file_hashes) or sum(all_sizes.values()) != int(manifest["source_bytes"]):
        raise RuntimeError("preflight payload set/size mismatch")

    output_dir = Path(os.path.abspath(str(args.output_dir)))
    target_assets = output_dir / "Assets"
    output_fd = open_or_create_directory_chain(output_dir)
    lock_name = ".robodojo-restore.lock"
    lock_fd: Optional[int] = None
    stage_root: Optional[Path] = None
    stage_recovery_path: Optional[Path] = None
    preserve_stage = False
    backup: Optional[Path] = None
    backup_name: Optional[str] = None

    try:
        try:
            lock_fd = os.open(
                lock_name,
                os.O_RDWR | os.O_CREAT | getattr(os, "O_NOFOLLOW", 0),
                0o600,
                dir_fd=output_fd,
            )
            lock_info = os.fstat(lock_fd)
            if not stat.S_ISREG(lock_info.st_mode):
                raise RuntimeError(f"restore lock is not a regular file in {output_dir}: {lock_name}")
            fcntl.flock(lock_fd, fcntl.LOCK_EX | fcntl.LOCK_NB)
        except BlockingIOError:
            raise RuntimeError(f"another restore process holds the lock in {output_dir}")

        initial_target = lstat_at(output_fd, "Assets")
        if initial_target is not None:
            if not args.force:
                raise RuntimeError(f"{target_assets} already exists; use --force for atomic replacement")
            if not stat.S_ISDIR(initial_target.st_mode):
                raise RuntimeError(
                    f"refusing --force because existing Assets is not a real directory: {target_assets}"
                )

        proc_fd_root = Path(f"/proc/self/fd/{output_fd}")
        if not proc_fd_root.is_dir():
            raise RuntimeError("Linux /proc/self/fd is required for pinned output directory access")
        stage_name = f".robodojo-restore-{uuid.uuid4().hex}"
        os.mkdir(stage_name, 0o700, dir_fd=output_fd)
        stage_root = proc_fd_root / stage_name
        stage_recovery_path = output_dir / stage_name
        shard_stage = stage_root / "shards"
        final_stage = stage_root / "final"
        os.mkdir(str(shard_stage), 0o700)
        os.mkdir(str(final_stage), 0o700)
        final_assets = final_stage / "Assets"
        os.mkdir(str(final_assets), 0o755)

        print(f"Extracting {len(shards)} isolated shards with {args.workers} workers...", flush=True)
        shard_roots = [shard_stage / f"{index:03d}" for index in range(len(shards))]
        with concurrent.futures.ThreadPoolExecutor(max_workers=min(args.workers, len(shards))) as pool:
            for shard, count in zip(shards, pool.map(extract_shard, shards, shard_roots)):
                print(f"  staged {Path(shard['archive_path']).name} ({count:,} files)", flush=True)

        print("Rechecking pinned archive descriptors after extraction...", flush=True)
        with concurrent.futures.ThreadPoolExecutor(max_workers=min(args.workers, len(shards))) as pool:
            observed_hashes = list(pool.map(lambda shard: sha256_fd(int(shard["archive_fd"])), shards))
        for shard, observed in zip(shards, observed_hashes):
            if observed != shard["sha256"]:
                raise RuntimeError(f"archive changed during restore: {shard['archive_path']}")

        print("Merging shards into a single verified staging tree...", flush=True)
        for name in sorted(file_hashes):
            parts = validate_payload_path(name)
            shard_index = path_to_shard[name]
            source = shard_roots[shard_index].joinpath(*parts)
            info = os.lstat(str(source))
            if not stat.S_ISREG(info.st_mode) or info.st_size != all_sizes[name]:
                raise RuntimeError(f"staged file type/size mismatch: {name}")
            parent = ensure_safe_parents(final_stage, parts[:-1])
            destination = parent / parts[-1]
            if lexists(destination):
                raise RuntimeError(f"merge destination already exists: {destination}")
            os.rename(str(source), str(destination))

        actual_paths = scan_tree(final_assets)
        expected_paths = sorted(file_hashes)
        if actual_paths != expected_paths:
            raise RuntimeError("restored path/type set differs from FILES.sha256")

        if not args.skip_file_check:
            print(f"Hashing {len(file_hashes):,} staged files...", flush=True)

            def verify_file(name: str) -> Optional[str]:
                path = final_stage.joinpath(*validate_payload_path(name))
                return None if sha256_regular_file(path) == file_hashes[name] else name

            failures: List[str] = []
            with concurrent.futures.ThreadPoolExecutor(max_workers=args.hash_workers) as pool:
                for failure in pool.map(verify_file, expected_paths):
                    if failure is not None:
                        failures.append(failure)
            if failures:
                raise RuntimeError(f"{len(failures)} staged files failed SHA-256: {failures[:20]}")

        current_target = lstat_at(output_fd, "Assets")
        if initial_target is None:
            if current_target is not None:
                raise RuntimeError("Assets appeared during restore; refusing to replace concurrent output")
            final_stage_fd = os.open(
                str(final_stage),
                os.O_RDONLY | getattr(os, "O_DIRECTORY", 0) | getattr(os, "O_NOFOLLOW", 0),
            )
            try:
                publish_new(final_stage_fd, output_fd)
            finally:
                os.close(final_stage_fd)
        else:
            if current_target is None or (
                current_target.st_dev,
                current_target.st_ino,
                current_target.st_mode,
            ) != (
                initial_target.st_dev,
                initial_target.st_ino,
                initial_target.st_mode,
            ):
                raise RuntimeError("existing Assets changed during restore; refusing atomic replacement")
            backup_name = f"Assets.backup.{uuid.uuid4().hex}"
            backup = output_dir / backup_name
            final_stage_fd = os.open(
                str(final_stage),
                os.O_RDONLY | getattr(os, "O_DIRECTORY", 0) | getattr(os, "O_NOFOLLOW", 0),
            )
            try:
                preserve_stage = True
                try:
                    publish_force(final_stage_fd, output_fd, backup_name)
                except PublicationRolledBack as error:
                    preserve_stage = False
                    raise error.original
                except OSError:
                    preserve_stage = False
                    raise
                except PublicationRecoveryError as error:
                    raise RuntimeError(
                        f"{error}; recovery staging preserved at {stage_recovery_path}"
                    ) from error
                preserve_stage = False
            finally:
                os.close(final_stage_fd)
    finally:
        if stage_root is not None and preserve_stage:
            print(f"Recovery staging preserved at {stage_recovery_path}", file=sys.stderr, flush=True)
        if stage_root is not None and not preserve_stage:
            shutil.rmtree(str(stage_root), ignore_errors=True)
        if lock_fd is not None:
            os.close(lock_fd)
        os.close(output_fd)
        for shard in shards:
            try:
                os.close(int(shard["archive_fd"]))
            except OSError:
                pass

    if backup is not None:
        print(f"Previous Assets preserved at {backup}", flush=True)
    if args.skip_file_check:
        print("RESTORE COMPLETE; STRUCTURE VERIFIED; CONTENT HASH CHECK SKIPPED", flush=True)
    else:
        print("RESTORE VERIFIED", flush=True)
    return 0


if __name__ == "__main__":
    raise SystemExit(main())