[artiz]

An HTML conversion kept no trace of the elements its items were made from: TreeItem has prov (page geometry) and source (a track segment), but nothing names the original <p>, its id, or where in the document it stood, so two paragraphs with the same text were indistinguishable and a caller keyed to the source's blocks could not tie a converted item back (#903). And a block's text could only be read through a serializer, with the serializer's separators in it: export_to_text() keeps the blank line that follows a paragraph, so the slice between two paragraphs is the paragraph plus "\n\n" (#904).

The item tree now carries the source's blocks (ItemTree::blocks): every element that produced an item, every non-inline element with text of its own (text nodes reached through inline elements only), and every empty <p>, heading, <li>, <dt>/<dd>, cell, <pre> or <caption>, in document order, as a SourceBlock — tag, id, class, the element path (html/body[1]/div[3]/p[1]: each step the tag and the 1-based position among the parent's same-tag children, so an anonymous paragraph is located too), the block's own whitespace-normalized text, and the items made from it. TreeItem::block points every item — an inline group's runs, a table's rich-cell groups included — back at its block. A block that produced no item says why (BlockOmission): hidden (display: none, <script>), empty, flattened into a table's grid (with the table item), merged into an enclosing block's item (a <p> inside a <li>). Every text node of the source belongs to exactly one block, so the blocks' texts together are the document's text. ItemTree::source_text([i, j]) reads selected blocks with one \n between them and nothing else — the contract #904 asks for; the default Markdown and plain-text serializations are untouched.

The walk records the element each item is created for (the flushed buffer's element, the <li>, the <img>, the <table>, …) and a linear post-pass over the DOM enumerates the blocks. On the 7 MB EUR-Lex consolidated text the issues were measured on (14 519 items, 15 626 blocks) the pass adds about 0.8 s to a debug-build conversion of 7.9 s. Rust API only: the JSON export is docling's schema, which has no slot for it, and stays byte-identical (the regression suite and the HTML conformance tests pass unchanged).

Tests: the issues' two reductions (the located <p class="norm"> and its 84-byte reference text; the two paragraphs joined by one newline), an inline group's runs, every omission reason, and the IAS 12 excerpt of that EUR-Lex text (tests/fixtures/html_blocks/eurlex_ias12_excerpt.xhtml: headings with ids, anonymous paragraphs, a flattened table, a hidden and an empty paragraph) — the coverage invariant, paragraph 79 by its locator with the reference text, and the export_to_text() slice the issue reports, unchanged.

Closes #903
Closes #904

Claude-Session: https://claude.ai/code/session_01ND9Uzdq6iouWJivti5dX8T