bugfix: multi-line block quote inside list item (bd-vet6)

[cscheid]

Summary

The tree-sitter qmd parser failed on a Pandoc-valid CommonMark construct: a list item containing a block quote whose paragraph spans multiple lines using the > continuation marker (e.g. - > a\n > b\n). Pandoc parses this as BulletList[[BlockQuote[Para[Str a, SoftBreak, Str b]]]]; we returned a parse error at line 2 col 6. The bug fires for every list marker (-, *, +, 1., 1), etc.), the 3+-line case, and any variant with content following the block quote — the original 2-line variant only escapes because the scanner's EOF path runs before the buggy code (pampa hides this by auto-appending a \n, so users always hit it).

Root cause is in crates/tree-sitter-qmd/tree-sitter-markdown/src/scanner.c. After a SOFT_LINE_ENDING, the scanner sets STATE_MATCHING | STATE_WAS_SOFT_LINE_BREAK. When the parser then asks for the next token at the trailing \n of the second block-quote-marked line, the STATE_MATCHING block at line 2040 calls match_line, which routes LIST_ITEM through its case 2 "blank-line continuation" branch — that branch advances past the \n. By the time control reaches the line-ending gate at line 2233 (which checks lookahead == '\n'), the newline is gone, so the gate skips. The scan returns false, tree-sitter retries with a different lex-state and gets _close_block which has no shift at the current parse state, and the parse errors out. The BLOCK_QUOTE match has no analogous \n-consuming branch, which is why nested block quotes (> > a\n> > b\n) work fine.

The fix bypasses the STATE_MATCHING block when STATE_WAS_SOFT_LINE_BREAK is set AND lookahead is \n/\r — the soft-line-break already accounted for the continuation prefix, so re-running match_line against the trailing newline of the same logical line is wrong; the line-ending gate handles it cleanly. Full investigation, neighborhood characterization, fix proposal, and end-to-end verification are in claude-notes/plans/2026-05-11-bq-multiline-in-list-item.md. The branch is structured as five commits, one per phase (failing tests → characterization → fix → e2e verification → close-out), so the reviewer can step through the work in order.

Test plan

• tree-sitter test: 476/476 (6 new corpus tests for marker variants + 3-line case)
• cargo nextest run -p pampa: 3685/3685 (7 new pandoc-match fixtures, all match Pandoc's native AST)
• cargo nextest run --workspace: 8804/8804
• cargo xtask verify (full, 9/9 steps including hub-client WASM build + tests): passed
• Manual end-to-end on the original reporter file and all marker / multi-line / follow-up content variants; pampa output matches pandoc -t native byte-for-structure
• Regression sanity checks: bq-in-bq, blank-line-separated paragraphs in list items, nested lists, and lazy continuation all still produce the same trees as before
• No .snap files touched

Closes bd-vet6.