feat(html): HTML accessibility-tree backend

[bdelanghe]

Summary

• Adds src/html.rs — a new language backend that projects HTML DOM into an accessibility tree (Vec<Def>) for semantic identity tracking
• Wires HTML into the existing dispatch: clean/smudge filter, git-ast inspect, git-ast match, and git-ast setup (registers *.html/*.htm gitattributes)
• Downgraded dep from tree-sitter-html = "0.23" to "0.20" to match the API style of tree-sitter-rust = "0.21" (both use a language() fn rather than the 0.23 LANGUAGE const)

How it works:

• canonicalize() — normalises attribute ordering (alphabetical), double-quotes all values, lowercases tag names; output is valid HTML (not JSON) so working trees stay readable
• inspect() — returns a Def for each semantically meaningful element (landmarks, headings, interactive, img). Non-semantic divs/spans are skipped
• content_hash — FNV-1a over role\0name\0tag\0sorted-k=v (role + accessible name + structure)
• shape_hash — same preimage with accessible name AND its source attrs (aria-label, title, alt) replaced by _, so two navs that differ only in label share a shape hash and fuzzy-match correctly
• subtree_hashes — one hash per semantic unit, feeding directly into the existing identity::match_defs Dice-similarity matching

Key implementation note: tree-sitter-html 0.20 represents void elements (<img />) as element → self_closing_tag (not self_closing_element), so element_to_def handles both start_tag and self_closing_tag children defensively.

Test plan

• 20 new unit tests in html::tests covering: canonicalization, attribute sorting, idempotency, DOCTYPE normalization, accessible name resolution (aria-label / alt / text / title priority), role mapping, implicit/explicit roles, inspect element discovery, content_hash stability, shape_hash rename detection, subtree_hashes presence
• All 97 tests pass (cargo test --lib)
• Full test suite including cucumber acceptance tests passes (cargo test)
• cargo build compiles clean (zero warnings)

PR checklist

1. Independent PR — HTML backend only; no changes to existing Rust/JSON paths
2. Changed codepaths verified — 20 HTML-specific unit tests + 5 cucumber integration tests pass
3. Root cause identified — N/A (new feature)
4. No duplication — FNV-1a copied from printer.rs (private there); a follow-up can extract it if the pattern repeats again
5. No unrelated changes — only Cargo.toml/lock (new dep), 6 wiring files, and the new src/html.rs

🤖 Generated with Claude Code