Tree-sitter migration — tracked spike

[razakadam74]

Follow-up to #66. Spike ran; this issue parks the migration so it stays live without committing engineering until evidence shows up.

Status

Parked. Feasibility proved, necessity not. Numbers and gotchas: #66 spike report.

Triggers to activate

Any one flips this from "tracked" to "do it":

• Detect signature changes of exported functions as breaking changes #61 (signature-change detection) needs full ASTs regex can't deliver
• 2+ wonky-message reports trace back to extractor limits
• a new-language regex ladder grows past ~60 lines

Constraints to preserve

• zero runtime deps in the published package
• <100 ms cold start
• single-binary build via bun build --compile still works

Open questions

• Native vs WASM — web-tree-sitter clears the constraints (cold 23.8 ms, ~2.3 MB per grammar). Native bindings break single-binary. WASM is the front-runner.
• Hunk vs full file — today gitmsg sees diff hunks; tree-sitter wants full files. Either git show both sides → parse → diff symbol tables (simpler, two extra git calls per file), or invent a hunk-friendly approach.
• Per-language order — TS first (already proven, biggest base), then Python, Go, C#. Keep regex as fallback for ungrammared languages.
• Build pipeline — vendor pre-built .wasm, or rebuild grammars in CI against a pinned tree-sitter-cli? Spike showed vendored third-party .wasm is fragile when external C scanners are involved.

What's validated

• tree-sitter-typescript via tree-sitter-wasms@0.1.13 parses real source cleanly
• 5-line query pulls only exported symbols — exported is a built-in tree property (no per-language visibility heuristic, unlike current regex extractors)
• cold path 23.8 ms · warm parse 0.42 ms · full src/ stress 3.15 ms/file

What's not validated

• the git show + symbol-table-diff approach end-to-end
• rebuilding grammars from source against a pinned runtime
• bundle size impact on bun build --compile

Non-goals

• not blocking 1.0
• not removing regex extractors when this lands — they stay as fallback
• not migrating all languages in one PR