[jelveh]

Follow-up to #139, stacked on it (base: docs-latency/a-docx). This one speeds up reading a .docx when it opens.

Results (Chrome, same 28-page, 1,500-paragraph test document as #139)

	#139	this PR
DocxPackage.open (median of 9)	505 ms	325 ms
Open → editable (app)	~1.20 s	~1.00 s
Autosave stall	87 ms	88 ms

The JS reading work dropped from ~315 ms to ~135 ms. What remains is mostly native XML parsing and inflate.

Causes and fixes

• Element-children lookups. Every child()/children() call copied the live children collection into an array, tens of thousands of times per open. They now walk siblings through helpers in docx-xml.js that keep the same results and exceptions, used across the reader modules.
• Relationships parsed repeatedly. The main document's relationships part (60 KB with 300 hyperlinks) was parsed seven times per open. The colors, hyperlinks, mail-merge and page-stories readers now share the per-bytes cache from Documents: instant typing and faster opening in long documents #139, and each hyperlink looks its relationship up by Id instead of scanning them all.
• Numbering:
  • Each level read its paragraph properties twice; now once.
  • The test document has 1,584 levels but only 36 distinct ones. Each distinct level's markup is now read once, and every level still gets its own copy.
• Paragraph source serialized on demand. Opening serialized each body paragraph's source XML with its inherited namespace declarations, about 3 MB of text, which most sessions never use. Opening now copies the paragraph and its declarations, and serializes the copy when it is first needed. Documents: instant typing and faster opening in long documents #139's idle-time save preparation spreads that out, so the first autosave doesn't stall. This is a separate commit that can be dropped on its own.
• serializeFragment indexes attributes instead of copying them.

Verification

• Opening and saving: the opened package model, the saved parts and an edited save are identical for all 2,257 fixture and corpus documents in Chrome.
• Saving on top of Documents: instant typing and faster opening in long documents #139: checked again with this branch: every document edited and saved twice in Chrome, byte-for-byte identical to main.
• Unit suite: 10,569 tests pass. The touched tests pass on Node 22. The build passes.
• New test docx-read-reuse.test.js, built from synthetic markup:
  • repeated numbering levels stay separate objects;
  • preview issues from a repeated level are still reported;
  • duplicate relationship Ids are still rejected;
  • unchanged paragraphs save byte-identically;
  • the deferred source is unaffected by later tree changes;
  • the child helpers match [...node.children].

Not done (would change behavior or need a broader audit)

• Skipping TextDocument's re-resolution of already-resolved paragraphs: one preview-recovered fixture differs.
• Native DecompressionStream instead of fflate: its errors on corrupt archives differ.

🤖 Generated with Claude Code