feat: add real documents that carry mathematics

ATTRIBUTIONS.md

@@ -142,3 +142,11 @@ Vendored WordPerfect-family test files. Per-file provenance (upstream path, revi
 - **Rendering tools:** the source graphs are ours; the SVGs and PDFs are what each renderer emitted from them. Graphviz 15.1.1 (`dot`/`neato`) is EPL-1.0, Mermaid CLI 11.16.0 is MIT, PlantUML 1.2026.0 is GPL, LibreOffice 26.2.5.2 is MPL-2.0, librsvg 2.62.3 (`rsvg-convert`) is LGPL-2.1-or-later, qpdf 12.3.2 is Apache-2.0, and Google Chrome 151 is proprietary. None of these licenses reaches the output: rendering a document is not a derivative work of the renderer, and PlantUML says so explicitly in its FAQ. No tool binary, jar or stylesheet is redistributed here.
 - **Embedded fonts:** a PDF carries subsetted glyph outlines of whatever fonts the renderer reached for, so the fixtures under `diagrams/pdf/` do redistribute font data. Each family was checked against the OS/2 `fsType` bit of the system font it came from, which is where a font states its own embedding terms: Helvetica and Liberation Sans report 0 (installable, no restriction), Times New Roman and Trebuchet MS report 8 (editable embedding). All four permit it. `scripts/build_diagram_pdfs.py` holds that allowlist and fails a build that reaches for anything outside it. That check is why `graphviz_cjk` ships as SVG only: macOS's Songti reports `fsType` 2, restricted, which forbids embedding without the owner's permission, so no CJK PDF is published until an open-licensed CJK face is available to render one.
 - **Modifications:** every fixture whose producer states its own graph also ships as a `*_geometry.svg` variant with that metadata removed by `scripts/strip_svg_graph_metadata.py` — Graphviz writes the node ids and the full edge list into `<title>` elements and again into XML comments; Mermaid into `id="L_start_auth_0"`, again into `data-id`, and the edge's waypoints into `data-points`; PlantUML into `id="Read config-to-Open input"`, again into a comment, again into a `data-entity-1`/`data-entity-2` pair, and once more into a `<?plantuml-src?>` processing instruction carrying its whole deflated source; and LibreOffice by tagging each group with its shape kind, which across a whole file is the node/edge partition. A fixture that names its own answer cannot measure geometry recovery. Stripping removes no geometry: each variant renders to a byte-identical PNG. `nested_transforms.svg`, `icon_nodes.svg`, `mixed_page.svg`, `two_diagrams.svg` and the three `negative_*.svg` files are hand-written. Ground truth is the source graph restated by node label, so it is independent of any recogniser's numbering.
+
+## xberg math corpus
+
+- **Citation:** assembled for xberg from published documents that carry mathematics, one per notation and container the extractor supports.
+- **Source:** 123 upstream projects and publishers. Each document's source URL is in `MATH_PROVENANCE.md`.
+- **License:** per document. 136 are vendored under permissive terms, chiefly MIT (36), CC-BY-4.0 (25), Apache-2.0 (21), US public domain (13), BSD-3-Clause (12) and MPL-2.0 (5). Four state a dual `CC-BY OR GPL-3.0-or-later` licence and are taken under the CC-BY arm. 12 carry ShareAlike, GPL or GFDL terms and are reference only, recorded without their bytes.
+- **Used here:** 136 documents vendored, 12 reference.
+- **Modifications:** none. Each document is byte-identical to what its publisher serves, so a fixture measures the real file rather than a rewritten one. No ground truth is derived from them here.

LICENSES.md

@@ -44,4 +44,7 @@ non-commercial benchmark use non-commercial data.
 | [ds4sd/DocLayNet](https://huggingface.co/datasets/ds4sd/DocLayNet-v1.1) | CDLA-Permissive-1.0 | vendor |
 | [google-research-datasets/hiertext](https://github.com/google-research-datasets/hiertext) | CC-BY-SA-4.0 | reference (ShareAlike — not committed) |
 
-Vendored sources: 6. Reference-only sources: 1. WordPerfect corpus provenance: `wordperfect/PROVENANCE.md`.
+| math corpus (136 documents, ~70 upstream projects) | MIT / Apache-2.0 / CC-BY / CC0 / US-PD / MPL-2.0 | vendor |
+| math corpus, copyleft subset (12 documents) | CC-BY-SA / GPL / GFDL | reference (not committed) |
+
+Vendored sources: 7. Reference-only sources: 2. WordPerfect corpus provenance: `wordperfect/PROVENANCE.md`. Math corpus provenance, per document: `MATH_PROVENANCE.md`.