feat(parser): recognize bare < as Str token (bd-j9cf)

[cscheid]

Summary

• A literal < outside math/code/HTML used to produce a hard parse error in qmd (minimal trigger: 1 < 2). After this PR, a < that does not start a recognized HTML construct parses as Str "<".
• Conservative scope: the scanner still attempts the existing HTML constructs first and only falls back to a new LT_STR_LITERAL token when the scan loop reaches EOF without finding a closing > / }. <5>, <,>, etc. continue to lex as HTML_ELEMENT and emit Q-2-9 — no behavior change for those.
• Tracks bd-j9cf. Plan + decision log: claude-notes/plans/2026-05-18-bare-lt-as-str.md.

What changed

• scanner.c::parse_open_angle_brace — advances past <, calls mark_end at <+1 as a fallback, and emits LT_STR_LITERAL if the scan loop reaches EOF without finding a closing delimiter. Each successful HTML/autolink emit gained an explicit mark_end after its advance past > so the fallback end position is updated when a real HTML construct does match.
• grammar.js — declares $._pandoc_lt_str as an external token and adds it as a third choice inside pandoc_str so downstream consumers see a uniform Str shape.
• treesitter.rs — extends the pandoc_str handler to split leading ASCII whitespace into a separate Space inline (the block-level scanner chomps preceding indentation into the external-token range; same pattern as the existing html_element / autolink handlers). Leading-only is deliberate: \<space> backslash escapes match the \\. regex branch and rely on trailing whitespace to produce U+00A0 via process_backslash_escapes.

Generated / regen files

• parser.c and grammar.json regenerated by tree-sitter generate; tree-sitter build.
• _autogen-table.json regenerated by crates/pampa/scripts/build_error_table.ts — the new external token shifted LR state numbers, so the Merr-style state→error-code mapping needed a refresh. All error codes are unchanged; only the (state, sym) keys moved.

Tests

• crates/tree-sitter-qmd/tree-sitter-markdown/test/corpus/lt-as-str.txt — 7 corpus cases (4 new behavior + 3 regression for <b> / autolink / HTML comment).
• crates/pampa/tests/test_bare_lt_str.rs — 10 unit tests through readers::qmd::read (4 new behavior + 6 regression: html_element, autolink, comment, \<, math $1 < 2$, code span `1 < 2`).
• crates/pampa/tests/roundtrip_tests/qmd-json-qmd/bare_lt_{simple,eol,unclosed_tag}.qmd — auto-picked up by the existing roundtrip runner.

End-to-end verification

$ cargo run --quiet -p pampa --bin pampa -- /tmp/q2-lt-test-bd-j9cf/simple.qmd
[ Para [Str "1", Space, Str "<", Space, Str "2"] ]

$ cargo run --quiet -p pampa --bin pampa -- -t qmd /tmp/q2-lt-test-bd-j9cf/simple.qmd
1 \< 2

$ cargo run --quiet -p pampa --bin pampa -- -t html /tmp/q2-lt-test-bd-j9cf/simple.qmd
<p>1 &lt; 2</p>

The qmd writer already escaped < as \< defensively (crates/pampa/src/writers/qmd.rs:1414), so round-trip works for free: emit Str "<", write \<, re-read via the \\. branch of pandoc_str.

Test plan

• tree-sitter test — 508/508 passing (501 existing + 7 new)
• cargo nextest run -p pampa --no-fail-fast — 3742/3742 passing, 2 skipped
• cargo nextest run --workspace --no-fail-fast — 8998/8998 passing, 195 skipped
• cargo xtask verify --skip-hub-build --skip-hub-tests — all steps passed (Rust + lints + fmt + trace-viewer)
• cargo xtask verify --skip-rust-build --skip-rust-tests — all steps passed (hub-client build + WASM tests)
• E2E binary verification on 1 < 2, foo <, a <foo fixtures with native / qmd / html / json writers — outputs inspected manually

Out of scope / notes

• No snapshot files needed updating. The only behavior change beyond rescuing bare < is the regenerated _autogen-table.json (same error codes, new state keys).
• Symmetric handling of a bare > is unnecessary: > is already inside PANDOC_REGEX_STR's punctuation class ([>.,;!?]).

🤖 Generated with Claude Code