[nicoburns]

LLM Contributions: Code changes generated with Devin.

Prototype of replacing ICU4X with cleft (dfrg's no_std, zero-dependency Unicode crate) for text analysis, on top of linebender#898 (merged with main; the base branch harfrust-0.15-plus-main is exactly that merge, so this PR's diff is the swap alone). Not meant to merge as is — it exists to measure. Full write-up with the method and raw data: see the session attachments; summary below.

What changes:

• Line, word and grapheme segmentation, bidi classes and paired brackets, script and emoji flags come from cleft; canonical (de)composition for font fallback comes from harfrust::unicode (harfrust 0.15). No icu_* crate remains in the dependency graph; parley_data/parley_data_gen are deleted (the few sets cleft does not expose — Cc, Variation_Selector, Emoji_Modifier(_Base) — are fixed ranges in analysis.rs).
• CharInfo::script is a parlance::Script; DictionaryProvider (icu_provider) is replaced by cleft::Dictionaries (set_dictionaries on the builders); complex-scripts enables cleft/dictionaries. The external_dictionaries example is removed (cleft's .dafsa generator is not published yet).
• Runs with different line-break/word-break options restart the scan at the last opportunity before the seam (LineBreaks::with_preceding) instead of carrying one character back.
• contributes_to_shaping uses Default_Ignorable_Code_Point (minus the Hangul fillers HarfBuzz draws) in place of General_Category=Format.

Results (x86_64 Linux VM, details in the write-up):

• Binary size of tiny_skia_render, release, stripped: main 2,703 KB → Upgrade to read-fonts 0.46, skrifa 0.49, harfrust 0.15 linebender/parley#898 3,023 KB → prototype 2,974 KB; with complex-scripts: 6,529 KB → 6,849 KB → 5,665 KB (−1.2 MB vs Upgrade to read-fonts 0.46, skrifa 0.49, harfrust 0.15 linebender/parley#898).
• tango benchmarks (parley_bench + the document group from Add larger "document scale" benchmarks for each pipeline phase linebender/parley#793), build-phase groups, median: Upgrade to read-fonts 0.46, skrifa 0.49, harfrust 0.15 linebender/parley#898 vs main −16.8%, prototype vs Upgrade to read-fonts 0.46, skrifa 0.49, harfrust 0.15 linebender/parley#898 −7.8% (noise floor of a self-compare: −1.0%). Latin/Japanese −6…−10%, Arabic −1.5…−4%.
• perf breakdown of the build phase on the prototype: shaping 54–75%, font selection & fallback 12–26%, analysis 8–20% (Japanese analysis share 15.8% → 9.5% vs Upgrade to read-fonts 0.46, skrifa 0.49, harfrust 0.15 linebender/parley#898; Latin barely moves because the per-character property loop in analyze_text is now the largest analysis leaf).

Tests: all pass, parley_tests snapshots unchanged. Two test_analysis expectations that encoded ICU4X behaviour changed: line-break: loose before U+2010/U+2013 needs cleft's loose_hyphens to match ICU4X/CSS Text 3, and cleft's Thai/Lao dictionaries segment กก into two words where ICU4X did not (that test now checks only the \n handling it was written for).

Changelog

Changed

• Breaking change: text analysis uses cleft instead of ICU4X. CharInfo::script is now a parlance::Script, DictionaryProvider is replaced by cleft::Dictionaries, and parley_data/parley_data_gen are removed.

Link to Devin session: https://dioxus.staging.devinenterprise.com/sessions/7519feeb2a15465c9b5d89742fb9a85c
Open in Devin Desktop: https://dioxus.staging.devinenterprise.com/desktop/session/7519feeb2a15465c9b5d89742fb9a85c?variant=devin-insiders

[staging-devin-ai-integration]

I'll fix CI failures and address comments from users with write access that start with 'Devin'.

• Disable automatic comment, CI, and merge conflict monitoring