ADR-0006: Highlighting via syntect; intra-line diff via similar

• Status: Accepted
• Date: 2026-06-10

Context

Two rendering-quality features define how the diff looks: syntax highlighting and intra-line (word-level) change emphasis. Both have a performance dimension — they are the expensive part of rendering and must never block scrolling (quality bar: 250k-line diffs at 60 fps).

The tempting maximalist choice is tree-sitter: real parse trees, and a path to structural diffing. But grammars must be bundled per language (binary bloat), C compilation per grammar, and structural diffing is a product we explicitly deferred (difftastic owns it; BLUEPRINT §3).

Decision

• syntect (Sublime-grammar engine, as used by bat and delta) for syntax highlighting, with precompiled binary assets, default-features = false
  • default-fancy (pure-Rust regex engine — keeps the no-C-toolchain build for contributors and musl targets trivial).
• similar for intra-line diffing: word-granularity inline spans computed per hunk-line-pair, unicode-aware.
• Both run lazily: highlight/intra-line only what scrolls into view, cached per file, warmed by a background thread. The input loop never waits.

Consequences

• Same visual quality users already trust from bat/delta, ~200 languages free.
• Pure-Rust build keeps cargo install margin working everywhere without oniguruma headaches.
• Laziness makes the giant-diff budgets achievable; the criterion benchmarks (ADR-0010) hold the line.
• Cost: regex-based highlighting is approximate; occasional mis-highlights in exotic languages. Accepted — every comparable tool shares this.
• Cost: no structural diff. Deliberate scope decision; a future tree-sitter ADR would supersede this one only with benchmarks and a binary-size budget.

Alternatives considered

• tree-sitter — better fidelity + structural future, but binary bloat, per-grammar C builds, and it drags us toward a product we chose not to build.
• No highlighting in MVP — would torpedo the “screenshot quality” first impression that drives adoption of this category of tool.
• imara-diff for intra-line — faster on large inputs, but intra-line works on short strings where similar’s ergonomics and unicode segmentation win. (imara-diff remains the candidate if file-level diffing ever moves in-process.)