Measured on 2026-08-25 using the 32 .replay fixtures under ../rl-combined-parser/test-files/.

The old parser's Boxcars binding and the Rust CLI were compared on the fields the Rust CLI emits:

• date, version, map, and game type
• duration
• team scores and derived winner
• player names, platforms, online IDs, teams, and headline stats
• goal frames, scorers, and teams

All 32 fixtures parsed successfully with both implementations. All scoped fields matched exactly: 32/32 fixtures.

Each implementation was warmed up once, then run over all 32 fixtures for five measured rounds.

The Rust CLI was approximately 14.1x faster than the old parser binding. Batch mode was approximately 1.88x faster than launching the Rust CLI once per replay.

The old-parser timing calls its compiled Rust binding from one persistent Python process. The per-file Rust timing includes process startup for every replay, making that comparison representative of command-line usage but not a strict parser-core benchmark.

rl-replay --batch replay1.replay replay2.replay ... emits a JSON array. The batch result was compared element-for-element with the 32 outputs produced by running the single-file command separately; all outputs matched.

cargo test
cargo build --release
git diff --check

The Rust unit test suite passes with five tests.

Measured locally on Apple Silicon using the release build:

The release workflow currently packages the executable together with the README and benchmark report, so published archives will be modestly larger. Other target platforms may differ slightly because of linker and binary format differences.