Sizing file-format bloat fixes offline before touching the writer

When a binary container file is far larger than a competitor, first write a tiny parser that assigns every byte to a class (metadata, index, payload, per-block footers) and check the classes sum exactly to the file size. Then size each candidate fix offline: compress the extracted pieces with the intended codec/level, rebuild hypothetical sections (e.g. hierarchical names instead of flat full paths, single copy of duplicated entries), and report saved bytes per fix, both raw and after compression, since fixes overlap once compression is applied. Also time the codec at the chosen level on the writer and the open path; high zstd levels on large metadata can cost seconds while level 3 costs milliseconds. Separate fixes that are writer-only from those that need reader/spec changes; a decompress-needed section cannot live in a segment type that readers are allowed to skip.