v1.5.0

Added

• Add max_rewind_cache_size to AsyncGzipBinaryFile and AsyncGzipTextFile. Non-seekable read streams now retain at most 128 MiB of compressed input by default for backward-seek replay; pass a byte limit to tune the cap or None to preserve the previous unbounded cache behavior.

Fixed

• Ensure AsyncGzipTextFile.close() does not finalize partially read decoder state. Closing after a partial multibyte read no longer raises UnicodeDecodeError or skips closing the underlying binary stream.
• Ensure AsyncGzipBinaryFile.close() still closes owned or closefd=True file objects when final compressor flush or trailer writes fail.
• Reject embedded NUL bytes in original_filename, which previously produced malformed gzip headers and unreadable archives.
• Reset max_decompressed_size accounting when a reader rewinds, so re-reading an otherwise under-cap archive after seek(0) no longer trips the cap.
• Treat file objects that report seekable() == False as non-seekable even if they expose a seek() method, using replay caching instead of calling a failing seek.

Changed

• Keep underscore-prefixed implementation helpers out of the top-level aiogzip.__all__; import internals from aiogzip._common, aiogzip._binary, or aiogzip._text only for unsupported internal testing/debugging.
• Broaden fileobj constructor type annotations to accept read-only objects in read mode and write-only objects in write mode.
• Use asyncio.get_running_loop() for executor dispatch in zlib offload paths.

Documentation

• Document bounded rewind-cache behavior for non-seekable streams and clarify the 128 MiB chunk_size cap.

Tooling

• Run the Python 3.8 syntax-compatibility guard in CI, not just pre-commit.
• Run twine check dist/* in the publish workflow before uploading artifacts.