The full corpus as byte-addressable release assets.
Individually gzipped books inside uncompressed tars: the container stays uncompressed so byte offsets remain exact, while each book is compressed separately (~2.6x measured). One HTTP Range request retrieves exactly one book — fetch ~170KB instead of ~450KB, with no dependency on anything else in the asset.
Every book carries a SHA-256 of its raw text in locator.sqlite, so upstream re-transcription is detectable rather than silently inherited. This is a pinned edition, not a mirror.
Offsets and integrity hashes: see locator.sqlite.gz.