pampa: record byte-for-byte provenance of Math.text (Math.text_source)

[cscheid]

Prerequisite for the quarto-math crate (bd-entbg6x3, plan claude-notes/plans/2026-09-21-quarto-math-and-native-docx.md, Phase 0). Strand: bd-ieldbghj.

Opened as a draft on purpose: the branch is complete and verified, but we are holding the merge for a while (it will follow #704). Expected rebase conflicts against main once #704 lands are three regenerable artifacts only: Cargo.lock, crates/pampa/snapshots/json/math-with-attr.snap, ts-packages/annotated-qmd/examples/academic-paper.json.

Why

Math.source_info spans the whole $…$ / $$…$$ node, but Math.text is not a plain substring of it once a math node spans lines: the reader folds each soft break of inline math to a literal \n and strips the block-continuation gutter (> , list indentation) from the interior lines of display math. A consumer that parses text further (the coming quarto-math crate) needs to map any byte of it back to the .qmd.

What

• quarto-pandoc-types: Math.text_source: Option<SourceInfo>, length equal to text.len(), map_offset(i) lands on the byte text[i] came from. None means no mapping finer than source_info is known.
• pampa reader: location::provenance_builder_with_context (a ProvenanceBuilder rooted the way node_source_info_with_context roots node spans). Inline math: text segments verbatim, a bare \n soft break verbatim, any other soft break (gutter, CRLF) a replacement so the folded newline maps to the source line ending. Display math: first line and every \n verbatim, each stripped gutter a zero-length replacement. The strip behavior itself is unchanged (bd-q6ed / bd-qpa2 still own it).
• JSON wire: new textS sidecar (bare pool ref) after t, emitted by both the Value and the streaming writer, read back by the reader; absent/null reads as None. Documented next to a/targetS in claude-notes/designs/raw-json-format.md.
• Filters and transforms: Lua assignment to .text clears the mapping; pandoc.Math() builds with None; crossref_render keeps the mapping when appending \tag{N} by concatenating a zero-width synthesized piece; every other constructor passes None.
• Strip lists: textS added to Rust remove_location_fields and the TS stripSourceInfoFields (+ unit test), pampaOracle, tiptap spike canonical/pampa; textS?: number on the TS Annotated_Inline_Math type.

Tests

• crates/pampa/tests/integration/math_text_source.rs (15): single- and multi-line inline math, blockquote and list gutters, display math plain / quoted / nested-quoted / listed / labeled / inside strong, several nodes per paragraph, JSON round trip of textS, JSON without the sidecar reads as None, and a sweep over every in-tree .qmd with math (21 files, 36 nodes) asserting each text byte maps inside the node span to the identical source byte unless it is a folded newline.
• quarto-core crossref_render: equation_tag_extends_text_source_instead_of_dropping_it.
• test_location_health validates text_source alongside source_info.

Snapshot changes (reviewed)

• crates/pampa/snapshots/json/math-with-attr.snap (1 file): each of the three Math nodes gains a textS ref and the pool gains the matching body ranges (e.g. 29..37 for E = mc^2 inside $E = mc^2$ at 28..38). Nothing else changed.
• ts-packages/annotated-qmd/examples/academic-paper.json regenerated with the live pampa -t json binary (guard test); only textS refs and pool rows added.

Verification

cargo nextest run --workspace (14,032 passed) and full cargo xtask verify (all steps passed) on this branch.

🤖 Generated with Claude Code

[posit-snyk-bot]

✅ Snyk checks have passed. No issues have been found so far.

Status	Scan Engine	Critical	High	Medium	Low	Total (0)
✅	Open Source Security	0	0	0	0	0 issues
✅	Licenses	0	0	0	0	0 issues

💻 Catch issues earlier using the plugins for VS Code, JetBrains IDEs, Visual Studio, and Eclipse.