[reneleonhardt]

Summary

• Fix Unicode data parsing and prevent cached files from mixing Unicode releases.
• Give canonical Hebrew forms equal widths. Make seven numerals correctly interrupt combining solidus overlays.
• Document supported Unicode rules and existing variation-selector, emoji, and control-character policies.

Changes

• Read property fields independently of comments and apply ordered @missing defaults before explicit values.
• Compare both width and scanner state in normalization tests.
• Reuse test buffers, remove debug output, and report missing benchmark data files.
• Run generator tests and build documentation in both feature configurations in CI.
• Add the default-off ascii-fast-path feature for ASCII-heavy workloads. CI tests it with and without CJK support.

Validation

• Focused Rust tests pass with CJK and the ASCII fast path independently enabled and disabled.
• Five generator tests, Clippy, formatting, and documentation builds with warnings as errors pass.
• Regenerated Unicode 18.0 tables and emoji test data match the committed files.
• Embedded ASCII, short Unicode, and mixed Unicode benchmarks pass with the ASCII feature enabled and disabled.

Performance

• Forced lookup inlining reduced time by about 8–20% on selected long Unicode samples without the ASCII shortcut.
• The measured scanner code grew by about 0.4–1.3 KB. Some tiny ASCII, CRLF, and mixed Unicode samples took about 0.2–0.7 ns longer.
• The final commit adds ascii-fast-path. The default build omits every ASCII shortcut check.
• With this feature, a 5,760-byte ASCII sample took about 20 times less time.
• Enabling it added about 0.5–1.5 ns on sampled short Unicode and CRLF inputs. A 63-byte mixed label took about 5% longer.
• Cargo combines features across dependencies. Another dependency can enable this feature for all callers. The iterator API bypasses the shortcut per call.
• These measurements depend on the compiler, input, and calling context.

The Old Korean extension was rejected because no net benefit for common workloads was proven.

[Manishearth]

This appears to be doing a ton of different changes with very little coherent story of what's going on. There's code cleanups mixed with docs changes mixed with behavior changes. Mixing behavior changes like this is just generally bad form in open source.

I don't have time to disentangle this. Separate well-scoped PRs might work, but also I don't really know what you're trying to do here.

If this was LLM done; please consider first talking to the maintainers about your desires and making sure we're on the same page. We have LLMs too; someone without any background on the crate proposing a pile of changes is strictly worse than us doing it ourselves.