FileDiskIo

A DiskIo backed by real files under basePath, via the platform RandomAccessStorage. Ports the piece↔file mapping logic of libtorrent's default_storage/mmap_storage (storage_utils.cpp): the torrent is a single concatenated byte stream, and a piece read/write fans out across whichever files its byte range overlaps.

Blocking file syscalls run on dispatcher (defaults to Dispatchers.Default; callers on JVM/Android should inject Dispatchers.IO, which is not available in commonMain).

Concurrency: a shared RandomAccessStorage does an internal seek then read/ write, so two concurrent peers hitting the same file handle would interleave a seek from one with the read from the other and corrupt data. Every handle is therefore guarded by its own Mutex making each seek+transfer atomic, and the handle map itself is guarded so lazy-open is race-free.

Part-file: blocks that belong to excluded (priority-0) files or BEP-47 padding files are not written to the real files. They are routed into a side <name>.parts file (libtorrent's part_file, part_file.cpp), keyed by piece index, so the data is still readable for hashing/verification and is reclaimed when the file's priority later rises. This mirrors libtorrent's behavior of never touching the destination file of a priority-0 file.

Allocation: storageMode chooses sparse files (the default) or full preallocation (ALLOCATE), matching libtorrent's storage_mode_t.

Resume note: checkExistingFiles always reports nothing as present, so every start looks like a fresh download. Verified resume hashes the on-disk pieces against the torrent hashes, and it runs in the engine's recheck path, where the expected hashes live.