Fix display math round-trip inside blockquotes / lists (issue #181)

[cscheid]

Summary

Fixes #181. Display math $$ ... $$ inside a blockquote retained the > continuation prefix verbatim in Math.text because pandoc_display_math in the tree-sitter grammar matches its body as a single regex token between $$ delimiters and so never goes through the block-continuation machinery. The qmd writer then correctly re-prefixed every line on output, doubling the > and adding another level on each round trip.

The fix is in the AST extractor (crates/pampa/src/pandoc/treesitter.rs):

• The opening $$ sits at column C = node.start_position().column. On every interior line of the math, bytes at columns 0..C are the accumulated continuation prefix added by all enclosing blocks (any combination of blockquotes, list items, fenced divs, etc.).
• Strip those bytes column-wise — but only if every one of them is in {>, space, tab}. Lines that don't fit that pattern (lazy continuation, where the user wrote no explicit > ) are left alone rather than having real content chewed off.

Because the column already encodes the cumulative prefix width, this handles arbitrary mixed nesting (> - $$, - > $$, > - > $$, > > $$, divs around either, etc.) without enumerating block types or walking ancestors per block kind.

I first attempted a structural grammar fix — making pandoc_display_math line-structured like pandoc_code_block so each interior line goes through _soft_line_break (which consumes block_continuation). That broke existing tests (Display math with list markers should remain a single paragraph, Display math inside fenced div should parse correctly) into ERROR nodes, because pandoc_display_math is an inline element living inside _inlines, and crossing soft line breaks at the inline level conflicts with _inlines's own line structure. The AST-extraction approach was the smaller, safer fix.

Triage doc and minimal repros live at claude-notes/issue-reports/181/ for context.

Test plan

Regression fixtures added under crates/pampa/tests/roundtrip_tests/qmd-json-qmd/:

• display_math_in_blockquote.qmd — the reporter's exact input
• display_math_in_nested_blockquote.qmd — > > $$ ... $$
• display_math_in_list_in_blockquote.qmd — > - $$ ... $$
• display_math_in_blockquote_in_list.qmd — - > $$ ... $$
• display_math_in_bq_list_bq.qmd — > - > $$ ... $$

All five previously diverged on qmd → JSON → qmd → JSON and now round-trip cleanly via test_qmd_roundtrip_consistency.

• cargo nextest run -p pampa — 3685 passed, 2 skipped
• cargo xtask verify --skip-hub-tests — full Rust workspace + WASM hub-client build + trace-viewer tests pass
• End-to-end check on the reporter's repro through cargo run --bin pampa: AST is clean (Math DisplayMath "\np = q\n"), round-tripped qmd is correctly > -prefixed, re-parsing yields the same AST as the original (idempotent)
• Hub-client vitest run not exercised — there is a pre-existing ERR_MODULE_NOT_FOUND failure on main HEAD unrelated to this change

Tracks beads bd-q6ed.

🤖 Generated with Claude Code