[Apeksha9389]

AsyncTranscriber.upload_file() calculates a file object's Content-Length from os.fstat(stream.fileno()). For standard-library gzip, bz2 and lzma readers, that descriptor describes the compressed file, while read() yields decoded bytes. The header therefore disagrees with the uploaded body. After a logical read offset, it can even become zero while audio remains.

Measure the stream's logical end and restore its position instead of using the backing descriptor. Run upload-body preparation on a worker thread because seeking a wrapped stream can perform decoding or disk I/O. Streams whose size cannot be obtained still use chunked transfer encoding. Caller-owned streams remain open.

Validation on Python 3.12: 748 unit tests pass, and Ruff check/format pass on the changed files. Six regressions using real gzip/bz2/lzma files, with and without an initial read, fail on unchanged master because the header differs from the body length. A further test verifies that a paused size probe leaves the event loop responsive. No provider requests were made.

Prepared with AI assistance; reproduced and tested locally.