Some new tricks

[hachikuji]

Approach:

• Dictionary compression: Repo names stored once with prefix compression; repos referenced by index
• Columnar layout: Separate streams for event types, repo indices, IDs, and timestamps
• Global sorting: Events sorted by (type, id) to minimize deltas
• Arithmetic coding: Repo indices encoded with frequency-based arithmetic coding
• Timestamp encoding: 2-bit categories (zero/one/small/large) pack common cases tightly
• Zstd finisher: Level 22 compression on the prepared binary stream

Down to 6.52MB. I tried a bunch of other stuff, but this was my limit. Can't wait to see how it gets beat!