yimingqiao hits 97.6% compression (5,093,619 bytes)

[YimingQiao]

This is an incremental improvement on top of kjcao’s codec. The gains come from three columns only:

• repo.name: split into owner/suffix streams and compress each with lpaq1, then reconstruct during decode
• event_id: rANS tuning (best precision kept)
• timestamps: structured delta/run + rANS encoding (kept in its best configuration)

Result

• yimingqiao size: 5,093,619 bytes (~97.6% compression)

What we tried that didn’t help

• repo_indices: tried rANS top‑k + ESC on the index stream; it lost the local/context patterns that lpaq1 captures, so size got worse.
• repo.name: tried sparse rename tables (event‑indexed renames) and explicit rename streams; renames are too rare, so metadata overhead dominated.
• repo.name (owner): tried an owner‑prefix dictionary + rANS prefix IDs; the dictionary + id stream cost more than lpaq1 on the raw owner list.
• repo.name (suffix): tried separating the suffix flags into a standalone rANS stream; it reduced context for the suffix payload and compressed worse overall.

[agavra]

Confirmed in CI/CD! Very impressive, we're so close to breaking 5MB it would be very cool if someone did.

┌────────────────────────┬────────────────┬────────────┐
│ Codec                  │           Size │ vs Naive   │
├────────────────────────┼────────────────┼────────────┤
│ Naive                  │    210,727,389 │   baseline │
│ yimingqiao             │      5,093,619 │     -97.6% │
└────────────────────────┴────────────────┴────────────┘