The Zip64 branch of the streaming header reader assumed a data descriptor always follows a >=4GB entry. When the entry instead has back-patched sizes (no descriptor), the following header was parsed as descriptor fields, overwriting the entry's correct size/CRC with central-directory 0xFFFFFFFF sentinels and, on non-seekable async streams, leaving the reader misaligned so the next entry threw.
Now it detects a header signature after the entry data and leaves the already-correct metadata untouched, parsing that header normally. The fix has been applied to both the sync and async readers.
Add a `tolerateTruncatedStream` option (default false) to `BZip2Stream.Create`/`CreateAsync`,
threaded through to `CBZip2InputStream`. When enabled, the decoder accepts a stream that has no
trailing footer - e.g. a truncated stream, or a sub-range of blocks extracted for random access:
- An end-of-input reached while reading a block/footer header (a true block boundary, tracked by
`expectingBlockStart`) is treated as a normal end of stream instead of throwing
`ArchiveOperationException("BZip2 compressed file ends unexpectedly")`. EOF in the middle of a
block, the block CRC, or the Huffman tables still throws.
- The whole-stream combined CRC in the footer is not verified, since a partial decode's running
combined CRC won't match the stored whole-stream value. Per-block CRCs are still enforced.
Default behaviour is unchanged (the flag defaults to false) and the change is applied symmetrically
to the sync and async read paths.
Tests (BZip2StreamTests):
- a footerless header-only stream decodes to empty with the flag and throws without it;
- a real block followed by end-of-input at the next block boundary decodes with the flag and throws
without it;
- a corrupted whole-stream combined CRC is tolerated with the flag and fatal without it;
- a complete, well-formed stream still round-trips with the flag set.
All existing BZip2 tests continue to pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Optimal is now a struct array, removing 4096 per-encoder object allocations.
- Encoder literal models are flattened into one pooled contiguous BitEncoder[].
- Decoder literal models are flattened into one non-pooled contiguous BitDecoder[].
- Updated async decoder paths to use the flattened model base index.
- Fixed LzBinTree pooled buffer nullability cleanup.
