docx-repair

name	docx-repair
description	Use to REPAIR a .docx damaged by a Google Docs or Word Online round-trip — the package/XML wiring, the footnote markup, leftover content controls, and heading styling. Triggers: 'Word won't open the docx / says it's corrupt', 'Google Docs export broken', 'fix the customXML error', 'recover unreadable content', 'phantom blank page', 'repair this docx'; AND 'footnotes broken after Google Docs', 'supra notes wrong after coauthor edits', 'cross-references point to the wrong footnote', 'bio footnotes show numbers instead of symbols (*, †, ‡)', 'author note shows 1 2 3 not star dagger', 'footnote numbering starts at the wrong number', 'separator line missing', 'doubled footnote marks (**, ††)'; AND 'boxes around text after Google Docs', 'content controls / doubled boxes around paragraphs', 'remove the boxes Word draws around headings', 'heading text isn't styled as a heading', 'headings look different / inconsistent heading formatting', 'blank/empty heading lines'; AND 'clean up Google Docs XML cruft', 'strip redundant run formatting', 'de-bloat docx after Google Docs', 'remove rsid bloat / no-op shading / explicit black' — or converting hardcoded 'supra note N' cross-references to auto-updating NOTEREF fields. Any OOXML-level repair on a .docx edited in a cloud editor, even if the user never says 'OOXML'. NOT for building a docx from markdown (law-review-docx) or exporting to PDF (docx-render).
user-invocable	false

DOCX Repair (Google Docs / Word Online damage)

Cloud editors damage a .docx in independent ways. This skill is the front door for all of them; run only the track(s) you need.

Damage class	Symptom	Fix
A. Package / OOXML wiring	Word pops "recover unreadable content?" or refuses to open; LibreOffice won't load; phantom blank page	scripts/docx_repair.py (plugin root) — §Package repair
B. Footnote & cross-reference markup	Bios show 1,2,3 not *,†,‡; numbering starts wrong; "supra note N" points to the wrong footnote; missing separator line	the footnote scripts — §Footnote repair
C. Document content (boxes + headings + cruft)	Visible boxes around freshly-edited text; heading-looking lines not styled as headings; same-style headings rendering differently; blank heading lines; bloated XML full of all-zero rsids, no-op shading, explicit b=0/i=0/u=none, redundant black color & default fonts	fix_footnotes.py's document.xml passes — §Content cleanup

They are decoupled: package repair fixes the part wiring (never touches content); footnote repair fixes the footnote markup; content cleanup strips Google-Docs leftover content controls and normalizes headings. A file can need any, all, or none. If you don't know which, run the package check first (it's a no-op on a clean package), then the footnote pass (it carries the content cleanup).

Heads-up: docx-render's Word path already composes docx_repair.py as a preflight, so a Google export "just renders" without a manual Track A. Run Track A manually when you need the repaired file itself (to hand back, edit, or footnote-fix), not just a PDF.

---

A. Package / OOXML wiring repair

Google Docs' .docx export emits OOXML that strict consumers reject while lenient ones (x2t) accept. Two recurring, concrete defects:

1. Case-mismatched OPC part references — the export spells the folder customXML (capital) in document.xml.rels and [Content_Types].xml but stores the part as customXml (lowercase). OPC part names are case-sensitive, so the part is unreferenced/untyped → Word says the document is corrupt.
2. <w:evenAndOddHeaders/> left in settings.xml → Word renders phantom blank pages.

scripts/docx_repair.py (at plugin root — it's a shared library doc_render.py composes, so it lives there, not in this skill's scripts/) fixes both, cheapest-first: case-normalize part references → drop the Word-breaking directive → if still structurally broken, reserialize via ONLYOFFICE docbuilder (clean OOXML, watermark-free, at the cost of a re-layout; opt out with --no-reserialize).

# CLI — detect + repair (in-place if dirty)
python3 "$SKILL_DIR/../../scripts/docx_repair.py" returned.docx fixed.docx
python3 "$SKILL_DIR/../../scripts/docx_repair.py" returned.docx --dry-run   # report only

import sys; sys.path.insert(0, "<plugin>/scripts")
from docx_repair import repair_docx, opc_integrity_issues
issues = opc_integrity_issues("returned.docx")   # [] = clean
repair_docx("returned.docx", "fixed.docx")        # -> RepairResult

Detail and the root-cause investigation: docs/investigations/2026-06-23_gdocs-customxml-case.md.

---

B. Footnote & cross-reference repair

Fix footnote formatting damage and convert hardcoded supra-note references to NOTEREF field codes.

Canonical Procedure — Google Docs round-trip

A law-review draft that round-trips through Google Docs every editing round (the OPV/Nadya case) comes back with both formatting damage AND stale cross-reference numbers. Run these in this exact order on the returned .docx. Use $SKILL_DIR for the absolute path to this skill's directory and keep the last-known-good draft as OLD.docx.

# 1. Accept tracked changes first (the coauthor's edits). Numbering and the
#    baseline remap both assume the FINAL accepted text. (Word: Review →
#    Accept All; or the document skill's accept-changes path.)

# 2. Repair footnote markup (separators, styles, bio custom marks, pStyles),
#    strip Google Docs content controls (the "boxes"), AND strip GDocs OOXML
#    cruft (hygiene pass, default-on). --normalize-headings restyles/cleans
#    heading paragraphs; --normalize-body-indent uniforms body indents (see §C).
uv run "$SKILL_DIR/scripts/fix_footnotes.py" returned.docx -o step2.docx \
  --normalize-headings --normalize-body-indent

# 3. Remap stale "supra note N" numbers against the known-good baseline, THEN
#    convert to NOTEREF fields. --baseline is what fixes the coauthor-shift.
uv run "$SKILL_DIR/scripts/create_crossrefs.py" \
  --docx step2.docx --output step3.docx --baseline OLD.docx

# 4. Render with WORD and eyeball it (LibreOffice numbers custom marks wrong).
uv run "$SKILL_DIR/../../scripts/doc_render.py" \
  step3.docx step3.pdf --renderer word --allow-word

Verify in the render: author bios show *, †, ‡ (not 1, 2, 3 and not doubled **, ††); the first real footnote is 1; a remapped reference (e.g. "Kahan & Rock, supra note 8") points to the correct footnote; no boxes around freshly-edited sections; same-level headings render identically. Then in Word, Ctrl+A, F9 to refresh the NOTEREF display numbers.

Procedure Facts (incident-grounded)

• Step 3 must run AFTER step 2. create_crossrefs --baseline counts footnote display numbers the way Word does — skipping the customMarkFollows bio marks. Until fix_footnotes converts the bios to custom marks, the bios are still numbered 1, 2, 3, so the alignment is off by the bio count and every remapped number is wrong. Running step 3 first silently mis-targets the cross-references it was supposed to fix.
• --baseline is not optional for the Google Docs case. Google Docs flattens every NOTEREF field back to hardcoded "supra note N" text, frozen at the prior draft's numbering. After a coauthor inserts/deletes footnotes the offset is non-uniform, so create_crossrefs without --baseline bookmarks by current position and mis-targets ~90% of references — the exact failure this skill exists to prevent.
• The remap's flagged list is a human cite-check queue, not noise. --baseline prints ⚠ … could NOT be remapped for references whose footnote content did not align one-to-one (inserts, deletes, densely-similar citation clusters). These are left unchanged on purpose — guessing a target you cannot prove is how a wrong citation ships. Surface the flagged list to the user for manual verification; do not suppress it.
• Render with Word, never LibreOffice, to verify. LibreOffice renders customMarkFollows numbering wrong (verified 2026-06-10), so bios that are actually correct can look broken — leading you to "fix" something that was right. doc_render.py --renderer word (or x2t) is ground truth.
• The goog_rdk "boxes" are an on-screen Word artifact — they do NOT export to PDF. Word draws boundaries around nested content controls in its editor, but neither Word's nor x2t's PDF export renders them. So a "no boxes" PDF is NOT proof the controls were stripped — verify at the XML level (unzip -p file.docx word/document.xml | grep -c goog_rdk → 0). fix_footnotes.py reports the strip count; trust that over the render.

When footnote repair applies

Common footnote symptoms in .docx files round-tripped through Google Docs or Word Online:

• Missing footnote separator lines
• Stripped paragraph styles (pStyle) on footnote bodies
• Stripped style definitions (FNStyleBest etc.) — the pStyle reference points at an undefined style and Word silently falls back to Normal
• Author bio custom marks (*, †, ‡) replaced with numbers, or rendered doubled (**, ††) when Google Docs welded the literal mark onto adjacent text
• Footnote numbering starting at the wrong number (offset from customMarkFollows bio footnotes)
• "supra note N" cross-references pointing to the wrong footnote after a coauthor inserted/deleted footnotes (numbers frozen by a Google Docs NOTEREF flatten)
• TOC separator paragraphs that inflate to fill a whole page
• Hardcoded "supra note N" / "infra note N" references that need to become auto-updating NOTEREF fields

Footnote scripts — quick start

Scripts are in this skill's scripts/ directory. Use $SKILL_DIR below as a placeholder for the absolute path to this skill (the directory containing this SKILL.md). Each script carries PEP 723 inline metadata, so uv run script.py auto-installs lxml — no --with lxml needed.

For the full Google Docs round-trip, follow the Canonical Procedure above (it chains these in the required order). Individual scripts:

# Dry run (show what would change)
uv run "$SKILL_DIR/scripts/fix_footnotes.py" path/to/file.docx --dry-run

# Fix cloud-editor footnote damage
uv run "$SKILL_DIR/scripts/fix_footnotes.py" path/to/file.docx -o fixed.docx

# Convert cross-references, remapping stale numbers against a known-good baseline
uv run "$SKILL_DIR/scripts/create_crossrefs.py" \
  --docx fixed.docx --baseline OLD.docx

# Refresh stale NOTEREF cross-ref numbers after a coauthor inserted/moved
# footnotes in Word (render-based, ground-truth; needs x2t or LibreOffice)
"$SKILL_DIR/scripts/refresh_noteref_caches.py" path/to/file.docx --verify

Which script do I want?

• Footnotes look broken after a Google Docs / Word Online round-trip (missing separators, wrong styles, bios numbered or doubled, mark/number mix-ups) → fix_footnotes.py.
• The doc still has hardcoded "supra note 42" text that should become auto-updating fields → create_crossrefs.py (add --baseline OLD.docx when the numbers went stale through a Google Docs round-trip).
• The doc already uses NOTEREF fields but a coauthor inserted/moved/deleted footnotes in Word and the cross-reference numbers are now wrong → refresh_noteref_caches.py (this is the common "Nadya emailed back tracked edits and the numbering is off" case).

Footnote scripts (detail)

fix_footnotes.py

Detects and repairs OOXML footnote damage. Handles multiple sources. Idempotent.

Google Docs / Word Online round-trip damage:

• Missing separator/continuation footnotes (id=-1, 0)
• Custom mark restoration for author bio footnotes (*, †, ‡). Handles both round-trip flavors: Word Online (symbol kept in a separate run) and Google Docs (symbol run deleted and the literal mark welded onto the adjacent text — e.g. <w:t>* Nadya Malenko</w:t>). A position-based lxml pass forces the first --bio-footnotes references and their footnote bodies into the canonical custom-mark shape and strips the welded literal so the mark does not render doubled (**, ††). Idempotent.
• Footnote ID renumbering (shifted by missing system footnotes)
• Missing paragraph styles (adds configurable pStyle to all footnotes)
• Wrong paragraph styles — reassigns pStyle="FootnoteText" (the Google Docs default) to FNStyleBest on every footnote paragraph so the whole doc uses the canonical law-review style.
• Missing style definitions — restores FNStyleBest (and the basedOn/link styles it depends on) from the canonical law-review reference template when a round-trip stripped them from styles.xml. The template is the same writing-legal/templates/law_review_template.docx that law-review-docx's build_docx.py feeds to pandoc, so style definitions stay consistent.
• Mutated style definitions — when the FNStyleBest / FNStyleBestChar block survives the round-trip but picks up Google Docs hyperlink-renderer residue (link-blue underline color <w:u w:color="0077CC"/> or white paragraph shading), the whole block is replaced from the template.
• TOC separator paragraph inflation (shrinks to near-zero height)

Pandoc-citeproc wrap parens:

• Strips the (...) wrapper pandoc adds around mid-footnote bracketed citations while preserving author-written explanatory parentheticals (which lack the double-whitespace XML signature).

Document content cleanup (see §C):

• Content controls (the "boxes") — unwraps every <w:sdt> tagged goog_rdk* (Google Docs suggestion-mode markers, often 3-deep, that Word draws as boxes), replacing each with its <w:sdtContent> children. Keeps non-goog sdts (real controls, the TOC). Runs by default across all content parts (document, footnotes, comments, headers, footers — not just document.xml); idempotent.
• OOXML hygiene (de-cruft) — strips redundant off/default run formatting, all-zero rsids, no-op shading, black color, and default-font residue across all content parts (document, footnotes, comments, headers, footers). Keeps every "on" property. Runs by default (--no-hygiene to skip); idempotent.
• Heading normalization (--normalize-headings, opt-in) — styles unstyled heading-looking paragraphs and strips direct formatting off every heading; see §C.
• Body indent normalization (--normalize-body-indent, opt-in) — uniforms body first-line indents; see §C.

Flags:

• --output / -o: Output path (default: overwrite input)
• --dry-run: Show what would change without modifying
• --bio-footnotes N: Number of author bio footnotes (default: 3)
• --crossrefs: Chain to create_crossrefs.py after fixing
• --fix-numbering: Fix numbering offset from customMarkFollows bio footnotes (adds numRestart, updates NOTEREFs and supra references)
• --normalize-headings: Normalize headings (off by default) — style unstyled heading-looking paragraphs (2b) AND strip direct formatting + delete empty heading paragraphs (2a); restores Heading1–4 style defs from the template if missing. See §C.
• --no-hygiene: Skip the Google Docs OOXML hygiene pass. Hygiene is on by default — strips all-zero rsids, redundant off/default run formatting, no-op shading, black color, and default-font residue across content parts; keeps all "on" formatting. See §C Feature 3.
• --normalize-body-indent: Apply the document's dominant first-line indent to body paragraphs that lack it (editorial; off by default). See §C.
• --template PATH: Reference template (.docx) to restore missing footnote style definitions from (default: bundled writing-legal/templates/law_review_template.docx)

create_crossrefs.py

Converts hardcoded "supra note N" references to NOTEREF field codes that auto-update.

What it does:

• Finds all supra note <number> patterns in document body and footnotes
• Creates bookmark targets on referenced footnotes
• Replaces hardcoded numbers with NOTEREF _RefFN<id> \h field codes
• Preserves italic formatting on "supra"

--baseline OLD.docx — remap stale numbers before converting (the coauthor-shift fix):

When a coauthor edited the draft in Google Docs, every NOTEREF field came back as hardcoded "supra note N" text frozen at the prior draft's numbering. After footnotes are inserted/deleted the offset is non-uniform, so converting those numbers as-is mis-targets ~90% of references. --baseline aligns the baseline and current footnote sequences by letter-only content fingerprint (difflib) and rewrites each stale number to its current value before conversion. The remap is cross-run aware ("supra" is its own italic run; the number lives in a following run). References whose footnote content does not align one-to-one (inserts, deletes, densely-similar clusters) are flagged for human cite-check, never guessed. Run this after fix_footnotes.py (so bio custom marks are already excluded from the count).

Flags:

• --docx: Input DOCX (required)
• --output: Output path (default: overwrite input)
• --dry-run: Report changes without writing
• --baseline OLD.docx: Known-good prior draft whose "supra note N" numbering is correct; remaps stale numbers to current numbering by content-identity alignment

refresh_noteref_caches.py

Refreshes the cached numbers on existing NOTEREF cross-reference fields after footnotes were inserted/moved/deleted in Word. Use when cross-references already ARE fields (not hardcoded text) but their numbers went stale.

Why the naive approaches fail (and this script's method):

• The offset is not uniform — +N to everything is wrong.
• Computing numbering from document.xml order is wrong: the 3 customMarkFollows author-bio footnotes are not counted in the numeric sequence, and a tracked footnote move makes XML order diverge from rendered order.
• LibreOffice's inline cross-ref render lies — it always recomputes NOTEREF on load and excludes unaccepted tracked-inserted footnotes, so it shows xrefs ~2 low even though it numbers the page-bottom markers correctly.

So the script uses the rendered page-bottom footnote markers as ground truth: render → extract markers → fingerprint-match each footnote to its true marker (longest-common-prefix, one-to-one, most-distinctive first) → set every NOTEREF cache to its target's marker. It also repairs NOTEREF field codes left dangling by Word's 40-char bookmark-name truncation (_RefBib_...2024 → the real _RefBib_...20). It deliberately does not add updateFields (that re-triggers the buggy recompute). Verify with a changes-accepted render — once inserts are accepted every engine agrees and the inline xrefs render correctly.

Requires: ONLYOFFICE x2t (preferred; onlyoffice-x2t nix package) or LibreOffice (soffice) as fallback, plus pymupdf (auto-installed via the inline script deps; run the file directly, e.g. ./refresh_noteref_caches.py file.docx).

Flags:

• -o / --output: Output path (default: overwrite input)
• --dry-run: Report the cache changes without writing
• --verify: Also emit a changes-accepted *_ACCEPTED_preview.pdf proof
• --soffice PATH: Path to the LibreOffice binary, used only when x2t is not on PATH (auto-discovered if omitted)

Scope (intentional): refreshes numbers only. It does not do editorial retargeting (e.g. "this xref should point to notes 210–212 instead of its current target"). That is a human decision — move the bookmark / change the NOTEREF target first, then re-run this to refresh.

Footnote reference

See footnotes-reference.md for detailed technical reference covering:

1. Run-level editing gotchas (NBSP, cross-run matching, xml:space)
2. Cloud editor damage patterns (what gets destroyed and why)
3. Direct ZIP surgery patterns (bypassing Document libraries)

Footnote Numbering Offset Fix

When author bio footnotes use customMarkFollows (*, †, ‡), they consume auto-numbers 1–3, causing body footnotes to start at 4. Fix by adding numRestart=eachSect to settings.xml and updating NOTEREF cached values.

Requires: A section break between title page and body. Render PDF with Word or x2t (scripts/doc_render.py at plugin root) — both honor numRestart; LibreOffice does not (renders restart numbering wrong; verified 2026-06-10, x2t restarts at 1 per section where soffice numbers continuously).

See footnotes-reference.md § 4 for details, code patterns, and the critical rule: numRestart goes in settings.xml ONLY (not in sectPr — causes all-zeros).

C. Document content cleanup (boxes + headings + cruft)

Three content passes carried by fix_footnotes.py, for damage that is neither package wiring nor footnote markup. All are idempotent and ride along with the canonical procedure's step 2.

Feature 1 — strip Google Docs content controls (the "boxes") — DEFAULT ON

Google Docs wraps freshly-written / suggestion-mode content in <w:sdt> content controls tagged goog_rdk_<n>, often 3-deep. Word's editor draws the nested ones as visible boxes around the text. The pass unwraps every goog_rdk sdt — replacing it with the children of its <w:sdtContent>, looping until none remain — and keeps non-goog sdts (real form controls and the TOC's docPartObj wrapper). It runs automatically (no flag); a doc with no goog_rdk sdts is a no-op. Text, footnotes, comments, and tracked changes are preserved verbatim.

Runs across every content part, not just document.xml. goog_rdk sdts also live in footnotes.xml and comments.xml (and could in headers/footers), where they render as boxes in the footnote/comment area. The strip iterates the same content-part list as the hygiene pass (document, footnotes, comments, headers, footers). Verified on OPV: 137 (body) + 104 (footnotes) + 28 (comments) → 0 remaining, with footnote italics/small-caps/superscripts and footnote count intact. (Before v5.55.1 only document.xml was processed — footnote-area boxes survived.)

Feature 2 — heading normalization — OPT-IN (--normalize-headings)

Two parts, run in order (2b then 2a) so a newly-styled heading is also formatting-cleaned:

• 2b — style heading-looking paragraphs that aren't headings. Detects short standalone paragraphs that either lead with a section marker (I.→Heading1, A.→Heading2, 1.→Heading3, a.→Heading4, (a)→Heading5) or are entirely bold (→Heading1), and are currently unstyled (pStyle None/Normal). Guards (why it's safe to run on a correct doc): skips anything already Heading* / TOC* / Title; skips Table-of-Contents entries (inside a docPartObj sdt, or a trailing-page-number row with a hyperlink/tab leader); skips Abstract and similar front-matter labels.
• 2a — strip direct formatting off every heading. Google Docs bakes per-paragraph formatting into headings, so same-style headings render differently. For each Heading* paragraph this reduces <w:pPr> to only <w:pStyle> (+ <w:numPr> for a genuine list heading) and strips <w:rPr> from every run, then deletes empty heading paragraphs. Result: headings derive entirely from the style definition.
• Heading style defs — restores Heading1–Heading4 definitions from the law-review template if a round-trip stripped them (same add-only restore used for FNStyleBest).

Feature 3 — Google Docs OOXML hygiene (de-cruft) — DEFAULT ON (--no-hygiene to skip)

Google Docs bakes redundant direct formatting into every run. The hygiene pass strips it across all content parts (document, footnotes, comments, headers, footers — NOT styles.xml/numbering.xml, where an explicit "off" can intentionally override an inherited "on"). On the raw OPV draft it removed ~33,800 nodes/attrs with zero visual change. Rules:

• A. Strip unconditional no-ops inside <w:rPr>: explicit-off toggles (<w:b w:val="0"/>, i, bCs, iCs, strike, dstrike, smallCaps, caps, emboss, imprint, outline, vanish), <w:u w:val="none"/>, <w:vertAlign w:val="baseline"/>, <w:rtl w:val="0"/>; no-op <w:shd w:val="clear"> (auto fill/color) anywhere; all-zero rsid attributes (w:rsid*="00000000" — the GDocs signature; real Word rsids are non-zero).
• B. Strip only when it matches the default: <w:color w:val="000000"/> (black); <w:rFonts> that names only the default body font (read from docDefaults, e.g. Garamond). Emptied <w:rPr>/<w:pPr> are then removed.
• CRITICAL — keep every "on" property: bold/italic on (<w:b/>, w:val="1"), real underline, smallCaps=1, super/subscript, non-black color, non-default fonts (Symbol for the * glyph, intentional Arial/Times). Verified preserved: italic case names, small-caps citations, and footnote superscripts all survive byte-identically to a --no-hygiene run.

Body indent normalization (--normalize-body-indent, opt-in, editorial) — applies the document's dominant first-line indent (mode of <w:ind firstLine> on Normal/unstyled body paras; firstLine=360 in OPV) to body paragraphs that lack it. Front-matter guard: only unstyled paras >60 chars after the first Heading1 (title/abstract/TOC excluded). Logged, never silent.

Content-cleanup Facts (incident-grounded)

• A heading whose text starts with a literal marker (a. …, 1. …) must NOT keep a numPr. 2a keeps numPr only for a genuine list heading (number auto-generated, no literal marker in the text). On the OPV baseline, a. Pure Robo-Voting carried a leftover Google Docs numPr while its siblings did not — keeping it would render a doubled marker ("1. a. Pure Robo-Voting") and the three sibling headings would not match. Dropping it makes a./b./c. resolve to identical <w:pPr><w:pStyle w:val="Heading4"/></w:pPr>.
• 2b's regression case is "no false restyling," not "no change." On a doc whose body headings are already styled (the OPV deliverable), 2b must report zero restyles — its only unstyled "heading-looking" paragraphs are TOC entries and Abstract, all excluded. 2a still changes that doc (it normalizes the already-styled headings); only 2b is the no-false-positive guard. Confusing the two leads to "fixing" guards that were working.
• The goog_rdk boxes do not export to PDF — verify removal in the XML. See the Procedure Fact above: a clean PDF is not proof. Check every content part, not just the body: for p in document footnotes comments; do unzip -p f.docx word/$p.xml | grep -c goog_rdk; done → all 0.
• goog_rdk content controls live in footnotes.xml and comments.xml, not only document.xml. The v5.54.0 strip processed the body alone, so 104 footnote + 28 comment controls survived and rendered as boxes in the footnote area of the OPV regen (found in production). The strip now iterates the same content-part list as the hygiene pass. A part-scoped cleanup that hard-codes document.xml will silently miss footnote/comment damage — scope both passes to the shared is_hygiene_part list.
• Hygiene never touches styles.xml or numbering.xml. An explicit "off" toggle (<w:b w:val="0"/>) inside a style definition can intentionally override an inherited "on" from its basedOn parent — stripping it there changes rendering. In run-level direct formatting it is pure GDocs cruft. The pass is scoped to content parts only; the audit's count came almost entirely from there anyway.
• Hygiene strips only explicit-off / default values, never bare or "on". A bare <w:b/> means bold-ON; <w:vertAlign w:val="superscript"/> is a real footnote mark; <w:rFonts w:ascii="Symbol"/> carries the * glyph. Stripping by tag name (rather than value) would silently kill italic case names, small-caps citations, and superscripts — the exact content this pass must keep. The compressed .docx shrinks only modestly (cruft is repetitive, compresses well); the uncompressed XML shrinks dramatically — judge the win by node count.

Related (document skill group)

This skill owns the REPAIR stage — package wiring (Track A), footnote markup (Track B), and document-content cleanup (Track C) — for a .docx damaged by a cloud editor. Adjacent stages:

• Build a styled .docx from markdown → law-review-docx (its build_docx.py chains this skill's footnote repair + NOTEREF conversion after the pandoc build).
• Render to PDF/PNG → docx-render / scripts/doc_render.py (Word path composes Track A's docx_repair.py as a preflight automatically).
• Footnote repair lives only here. fix_footnotes.py is the single canonical Google-Docs / Word-Online footnote fixer; there is no second copy. (Bluebook's create_crossrefs.py + audit_crossref_targets.py remain a deliberate, actively-used cross-reference fork with their own retargeting strategy — a different concern, not a footnote-fix duplicate.)

See the full document skill group.