IR Diff Engine

Status: Public surface shipped (M2.5). DocxDiff (Docxodus/DocxDiff.cs) is the default comparison engine as of v8.0.0 (2026-07-29) — decision D4 is resolved. WmlComparer (Docxodus/WmlComparer.cs) is the older engine: still available, but feature-frozen — it will not gain new capabilities going forward. See CHANGELOG.md under [8.0.0] for the flip itself.

Engine selector (M-B): a shared ComparisonEngine selector (WmlComparer = 0, DocxDiff = 1) lets a caller pick the engine on the CLI (redline --engine=), WASM, and npm surfaces, routed through the single DocxCompare.Compare dispatch owner (Docxodus/DocxCompare.cs). WmlComparer keeps wire value 0 for backward compatibility, but an omitted selector now resolves to DocxDiff on every surface (CLI, WASM, npm/worker) — WmlComparer requires an explicit selector.

The IR diff engine is a structure-aware DOCX comparison engine built on Docxodus' intermediate document representation (IR). It is the write-side analogue of the read-only IR pipeline that backs the markdown projection: it reads two documents into anchor-addressed IR snapshots, computes an edit script between them, and renders that script three ways — native tracked-changes markup, a consumer revision list, or the script itself as JSON (diff-as-data).

What it produces

The NVCA model voting agreement against a reproducibly edited copy, compared with DocxDiff.Compare and rendered by WmlToHtmlConverter with RenderTrackedChanges: true. The screenshot fixture uses DocxSession for surgical text edits and Open XML block edits for the move/delete/insert. Four families of change appear in one frame, all recovered structurally:

In the screenshot	Markup emitted
(f) "Qualified Key Holder" struck in red	w:del
The "Sanctions Authority" definition underlined in green	w:ins
Cascaded labels such as struck (a) followed by inserted (b)	w:numPr/w:numberingChange rendered as a deleted/inserted marker pair
Series A for a blank, means for shall mean and include	w:ins/w:del at token granularity inside an otherwise Equal block
The interpretation clause struck at the bottom in purple, re-inserted at the top	w:moveFrom/w:moveTo, paired by w:name and surfaced as one MoveGroupId

Note the paired list labels: the old letter is struck and the new letter is inserted wherever a structural edit shifts the automatic counter. The paragraph text remains unchanged; only its resolved numbering marker is redlined.

It is a sibling to WmlComparer in the comparison family. The differences that motivate it:

1. Anchor-addressed revisions. Every revision carries the stable block anchor(s) (kind:scope:unid) it derives from — the same anchor grammar as the markdown projection and DocxSession. A revision can be located in the projection or fed straight to a DocxSession mutation. WmlComparer.WmlComparerRevision has no anchors.
2. Diff-as-data. The edit script serializes to stable JSON, so the diff is storable, transportable to non-.NET consumers, and auditable. WmlComparer only produces an OOXML document or an in-memory revision list.
3. A modeled IR. Comparison runs over the IR's typed blocks/runs/format records rather than raw atom streams, which makes table row/cell-precise diffs, footnote/endnote scope diffs, and modeled-format-change detection first-class.

Public surface

public static class DocxDiff (Docxodus/DocxDiff.cs):

Method	Returns	Purpose
Compare(left, right, settings?)	WmlDocument	Tracked-changes document with native w:ins/w:del/w:moveFrom/w:moveTo/w:rPrChange markup. Satisfies the WmlComparer contract: AcceptRevisions(result) ≡ right, RejectRevisions(result) ≡ left at the per-block text level.
GetRevisions(left, right, settings?)	IReadOnlyList<DocxDiffRevision>	The consumer revision list, rendered directly off the edit script (no produce-then-reparse round-trip).
GetEditScriptJson(left, right, settings?)	string	The edit script as indented JSON — the diff-as-data differentiator.
GetSemanticChanges(left, right, options?)	SemanticChangeSet	Stable, versioned verification schema covering content, formatting, structure, relationships, review data, media, and opaque package changes.
GetSemanticChangesJson(left, right, options?)	string	Deterministic docxodus.semantic-changes JSON; use the returned change set's ToCanonicalJson() when compact bytes are required for hashing/signing.

Supporting public types: DocxDiffSettings, DocxDiffRevision, DocxDiffRevisionType, DocxDiffFormatChange, DocxDiffRevisionGranularity, DocxDiffFormatComparison. All #nullable enable, fully XML-documented, no static or process-global state (multi-author / consolidate-compatible — author flows per call via DocxDiffSettings.AuthorForRevisions).

SemanticDiff is the audit-oriented sibling of these renderer surfaces. It keeps the edit script as its alignment authority, projects all modeled families into a durable schema, and supplements them with bounded package facts so relationship-only or unknown-part edits cannot disappear. See semantic_diff.md for its schema, suppression, transport, and performance contract.

Anchor grammar and DocxSession interop

DocxDiffRevision.LeftAnchor / RightAnchor are block anchors of the form kind:scope:unid — e.g. p:body:a1b2c3d4 (a body paragraph), li:body:… (list item), tbl:body:… (table), p:fn3:… (a paragraph in footnote 3). The kind/scope match the markdown projection's and DocxSession's anchor grammar, so:

• A revision resolves to a location in the markdown projection (review UIs, blame).
• A revision can be passed straight to a DocxSession call (ReplaceText, GetBlockMetadata, AddAnnotation, …) on the corresponding document.

A left anchor resolves against the left document's IR; a right anchor against right. Anchor presence by revision type: Inserted → right only; Deleted → left only; FormatChanged → both; Moved source → left, Moved destination → right. A token-level revision inside a modified/moved-and-edited block carries the enclosing block's anchor(s).

Pipeline

                    settings (DocxDiffSettings.ToIrDiffSettings → IrDiffSettings)
                         │
left  ─ IrReader.Read ──▶ IrDocument ─┐
                                      ├─▶ IrEditScriptBuilder.Build ─▶ IrEditScript ─┬─▶ IrMarkupRenderer.Render   ─▶ WmlDocument
right ─ IrReader.Read ──▶ IrDocument ─┘                                              ├─▶ IrRevisionRenderer.Render ─▶ revisions
                                                                                     ├─▶ IrEditScriptJson.Write    ─▶ JSON
                                                                                     └─▶ SemanticDiff projector   ─▶ versioned semantic JSON

Internal stages (all internal, under Docxodus/Ir/Diff/):

• IrReader — reads a WmlDocument to an IrDocument (anchor-indexed blocks; accepted-revision view; provenance off for the diff path). Shared with the markdown projection.
• IrDiffTokenizer — splits IR runs into word/separator/atomic tokens with match keys (case folding, NBSP conflation, hyperlink-target-in-key, field transparency). The diff's tokenization, NOT an IR fact.
• IrBlockAligner — unique-hash (ContentHash, FormatFingerprint) anchoring → LIS spine → in-order gap fill; relocations fall off the spine as moves; similarity-based in-gap pairing + cross-gap fuzzy moves. When a reorder admits several equal-length spines, a structural-anchor tie-break keeps heavy blocks (tables / section breaks / opaque) anchored and relocates the lighter paragraph instead (a max-WEIGHT LIS, cardinality-primary so the move count is unchanged; fires only when the plain LIS relocated a structural block, so the paragraph-only path is byte-identical). Beyond cleaner two-way markup, this stops an ambiguous paragraph move across a table boundary from spuriously moving the table — which in Consolidate would contest the whole table block and block per-cell composition of a second reviewer's disjoint table edit (issue #229).
  • The in-place ORDER invariant (EnforceInPlaceOrderMonotonicity, issue #288). EmitEntries walks the RIGHT document and emits each left-owning unit at its right position, so reject ≡ left is exact in ORDER — not merely as a word multiset — only when the left → right map is strictly increasing across every in-place unit: 1:1 pairs and split/merge groups. Each forming pass enforces order-preservation against the pairings that existed when it ran (InOrderRefine's crossing bounds, SameSlotPair/JunctionPair's maxBelow/minAbove sweeps, ReleaseCrossingModifiedPairs' post-normalization), but nothing re-checked what a LATER pass added — and the split/merge containment scan runs after the refinement passes. A verbatim-duplicate block never anchors (BuildUniqueIndex keys on content unique to each side), so which occurrence pairs with which is settled in-gap, and the losing occurrence can end up on the wrong side of a group formed afterwards. A single global pass therefore enforces the invariant once: keep a maximum-weight strictly-increasing subsequence of (left → right position) and release the rest to plain Deleted/Inserted (always reversible), weight = blocks kept paired, so an ambiguous cut sacrifices the pairing holding the least content. It runs BEFORE cross-gap move detection, so a released pair that IS a relocation comes straight back as a Moved — usually a better reading than the in-place pairing it replaced. Fast path: an already-monotone alignment returns after one linear scan, untouched. Pinned by IrAlignmentAsserts.AssertLeftOrderReconstructible (asserted on every aligner test) and by DocxDiffFuzzRoundTripTests' block-sequence comparison.
• IrTokenDiffer — Myers O(ND) token diff inside a paired block (Equal/Insert/Delete/FormatChanged).
• IrTableDiffer — nested table row/cell diffs (a cell-text edit surfaces as a token diff inside that cell, not a whole-table blob).
• IrEditScriptBuilder — assembles the IrEditScript from the alignment + token/table diffs, including footnote/endnote scope ops.
• IrMarkupRenderer / IrRevisionRenderer / IrEditScriptJson — the three renderers above.

Edit script

The IrEditScript is an ordered list of block operations (IrEditOpKind), plus a parallel noteOps list for footnote/endnote scopes:

Kind	Meaning	Anchors
EqualBlock	Both sides identical	both
FormatOnlyBlock	Text-equal, modeled format differs (w:rPrChange-grade)	both
ModifyBlock	Same block, edited (carries a nested token/table diff)	both
InsertBlock	Right-only block	right only
DeleteBlock	Left-only block	left only
MoveBlock	Relocated block (source + destination ops share a moveGroupId)	source: left, dest: right
MoveModifyBlock	Relocated AND edited (the case WmlComparer cannot express as a move)	source: left, dest: right
SplitBlock	One left paragraph split across N≥2 right paragraphs (M2.6)	left + splitMergeAnchors (the N rights)
MergeBlock	N≥2 left paragraphs fused into one right paragraph (M2.6)	right + splitMergeAnchors (the N lefts)

Edited-move markup boundary. A MoveModifyBlock remains rich in the edit-script and revisions surfaces: its destination carries the token diff describing the edit within the relocation. Produced OOXML deliberately does not nest that token diff inside the move range. It emits the complete left paragraph as w:moveFrom and the complete right paragraph as w:moveTo, paired by w:name. Word-class consumers process ordinary w:ins/w:del nested among destination w:moveTo fragments as independent revisions; Reject All can therefore restore a nested deletion after rejecting the destination move and leave an orphaned old word. Complete move halves make accept/reject atomic while preserving the detailed change in the non-markup APIs (issue #359).

Footnote/endnote structural fidelity. The noteOps carry per-note block diffs that the markup renderer applies inside the produced footnotes/endnotes part, after which IrMarkupRenderer.RenumberNoteIds re-sequences every reference + definition into body-reference document order (mirroring WmlComparer's ChangeFootnoteEndnoteReferencesToUniqueRange). The invariants this path guarantees — verified by DocxDiffScenarioTests.Scenario_PreservesFootnoteStructure (every edit×feature scenario) and DocxDiffFootnoteRobustnessTests, with a headless-LibreOffice load backstop:

• Definition ids are unique and every body reference resolves to exactly one definition. A continuationNotice (or any typed) reserved note at a positive id keeps its id and leads the part; real notes renumber to a range disjoint from the kept reserved ids (the counter seeds above the highest positive reserved id), so a reserved note never collides with a renumbered real note.
• A reference is never dropped. A footnote/endnote reference is a zero-width inline; the token-diff run rebuilder (SourceRunModel.Slice) attributes a boundary zero-width to exactly the op whose token range owns it (ZeroWidthBoundaries), so a reference at the tail of an edited paragraph survives and is never double-counted.
• A reference-deleted note's definition is emitted once. When an edit removes a reference but the definition lingers in both stores, the builder reconciles the orphan into a single matched pair (identity = w:id) rather than a delete + an insert of the same id; the renumber pass links a w:del reference to its preserved definition so it stays resolvable on reject even when its id ≠ its reference ordinal.
• A note-in-note reference is renumbered with its target. A footnote/endnote whose body cites another note keeps a reference inside its definition body; the body-only renumber walk never visits it, so RenumberNoteIds records each definition's old→new id and RemapNestedNoteReferences sweeps BOTH note parts once — after both kinds' passes — with both maps, so SAME-kind nesting (a footnote citing a footnote) AND CROSS-kind nesting (an endnote ref inside a footnote body, or vice versa) are both remapped and neither dangles on accept/reject. (Note-in-note references are valid OOXML but LibreOffice Writer's DOCX import cannot load any document that contains one, cross-kind or same-kind — so the round-trip oracle here is the structural reference-resolvability check and Word, not LibreOffice.)

Hyperlink structural fidelity (w:hyperlink). An edit that straddles a w:hyperlink anchor slices the link across multiple token ops, so SourceRunModel.Slice emits one w:hyperlink wrapper per contributing op (each stamped with the source link's document-order ordinal, pt:SourceLinkId, plus its resolved target pt:SourceLinkTarget). CoalesceAdjacentHyperlinks then rejoins a maximal run of same-ordinal fragments back into ONE w:hyperlink — so the output is byte-faithful to the source's single link — merging a run when EITHER it carries a plain (Equal) run (an intra-anchor text edit) OR every fragment resolves to the SAME target (issue #232: a single link whose entire anchor is replaced with no shared token is still one link with an unchanged href). The merge keys on the resolved target (external URI, or #anchor for an internal link — resolved via the part annotation on the retained source tree, exactly as IrReader does), not the r:id string: this is what keeps the WC019 whole-anchor retarget (text and href change) split into two links — its del/ins fragments share the same r:id string at coalesce time but resolve to different targets, so the pure w:del-link (old target) and w:ins-link (new target) must stay separate for the post-assembly r:id remap (ImportHyperlinkAndExternalRelationships). Genuinely distinct adjacent links that merely share a target carry different ordinals and never group. Proof: IrMarkupRendererTests.Hyperlink_* (internal edit, fully-replaced same-target → one link, whole-anchor retarget → two links, adjacent-distinct-same-target).

Comment fidelity (Covered). Comments (w:commentRangeStart/End/w:commentReference) are dropped from IR paragraphs (reader rule N15 records them as IrCommentStore char-offset spans for the markdown projection only), but they are carried through the fine token-diff path as AlwaysKeep zero-width markers — exactly like bookmarks — so an edited commented paragraph gets fine per-word w:ins/w:del markup with its comment anchors intact (no whole-block bail). Three passes guarantee integrity, the comment analogue of NormalizeBookmarks:

• MergeRightCommentDefinitions copies any RIGHT-only w:comment the emitted right-sourced content references into the LEFT-based comments part (creating it if the left had none), plus the referenced commentsExtended (w15:commentEx paraIdParent reply links) and commentsIds entries — so a right-added comment never dangles.
• NormalizeComments reconciles the body so every commentReference resolves to exactly one w:comment, every commentRangeStart is unique + pairs 1:1 with a commentRangeEnd, and an unchanged comment survives both accept and reject: (A) a common comment with a bare survivor collapses to a single bare range; (A2) a right-added / left-deleted comment's bare markers are wrapped in w:ins/w:del so they toggle with their side; (B) a wholly-rewritten comment's del/ins copies renumber the deleted copy to a fresh id + a cloned definition — the comment-dedup analogue of the bookmark renumber-collision — with its OWN fresh w14:paraId + a cloned commentsExtended entry so a reject-side threaded reply keeps its parent link; (C) orphaned markers are paired/dropped so the output is always schema-valid + fully resolvable.

DocxDiff is strictly ahead of the blessed WmlComparer oracle, which drops comments entirely on any edit. Verified by DocxDiffCommentStructureTests (9-shape synthetic corpus) + DocxDiffCommentRealDocTests (vendored TestFiles/DD/DD002-DenseComments.docx) under OpenXmlValidator schema validity, a comment-structure round-trip, and a headless-LibreOffice comment oracle (tools/diffharness/lo/lo_comment_check.py). See docs/ooxml_corner_cases.md (comment threading is keyed on w14:paraId; the validator's paraId-uniqueness blind spot).

Header/footer scopes (2026-07-03 campaign — strictly ahead of the oracle, like comments). HeaderFooterOps carries per-STORY block diffs for the header/footer scopes — mirroring Word Compare's default-on "Headers and footers" granularity, which WmlComparer lacks entirely (it ignores header/footer differences). Gated by DocxDiffSettings.CompareHeadersFooters (default true; false restores the carry-left-verbatim behavior exactly). Stories pair per (section ordinal × occurrence kind) over each side's EFFECTIVE story grid (explicit w:headerReference/w:footerReference, else Word's previous-section inheritance), de-duplicated to distinct part pairs — scope names (hdr1…) are positional per-document labels and never the pairing axis. A matched story runs the block aligner exactly like a note (token-level ModifyBlock diffs); an all-Equal pair emits nothing (unchanged stories keep the verbatim part carry-over, preserving the revisionsInInput pins); one-sided stories are whole insert/delete. The markup renderer rebuilds a changed story inside its part via the shared block-op dispatch (shared revision-id counter; per-part media/hyperlink rel import — rels are part-scoped), creates inserted stories as fresh parts + references (ensuring w:titlePg for First, w:evenAndOddHeaders for Even), and marks deleted stories' content w:del (part + reference stay; accept leaves an empty story — Word's own behavior, empty ≡ absent at the text level). Fine-granularity revisions include hdr/ftr-anchored entries appended after note revisions; WmlComparerCompatible excludes them (its contract is the oracle's revision set, which has none — this keeps the 179-count parity scoreboard meaningful by construction; the differential harness likewise filters hdr/ftr revisions with the oracle-cannot-produce rationale). v1 ceilings: sections pair by ordinal; sectPr/settings visibility flags are ensured, not w:sectPrChange-tracked; unreferenced parts aren't compared; Consolidate does not merge header/footer scopes (see v1 limitations). Proof: IrHeaderFooterDiffTests, the renderer battery's story round-trip (AssertRoundTrip), and DocxDiffHeaderFooterSmokeTests (synthetic battery + the WC004 real pair + a headless-LibreOffice render oracle, tools/diffharness/lo/lo_headerfooter_check.py).

A ModifyBlock over a paragraph carries a tokenDiff; over a table, a tableDiff (row ops with nested cell ops); a textbox-bearing block carries textboxDiffs. A SplitBlock/MergeBlock carries splitMergeAnchors (the plural side, document order) plus segmentDiffs — one COMPLETE token diff per member, whose singular-side spans are slice-local; the slices tile the singular side's token stream exactly, in order (the partition invariant: slice i's length = Σ of segment i's non-Insert left-span lengths for a split, non-Delete right-span lengths for a merge — boundaries are implicit in the diff ops, never stored). N:M (plural on both sides) is physically representable by the nullable fields but rejected by the test-side AssertSplitMergePairing (a SplitBlock must carry a null rightAnchor, a MergeBlock a null leftAnchor, no anchor may appear in two ops' splitMergeAnchors) and never emitted by the builder — the pairing assert is the load-bearing scope ceiling. The JSON is a faithful serialization of this structure (top-level operations + optional noteOps + optional headerFooterOps; splitMergeAnchors/segmentDiffs appear only on split/merge ops and each optional array only when populated, so earlier scripts serialize byte-identically), and is deterministic for identical inputs.

Settings

DocxDiffSettings is the public mirror of the internal IrDiffSettings; it exposes the consumer-relevant subset and maps onto it in ToIrDiffSettings().

Public setting	Default	Maps to	Notes
AuthorForRevisions	"Open-Xml-PowerTools"	IrDiffSettings.AuthorForRevisions	matches WmlComparerSettings
Deterministic	true	IrDiffSettings.Deterministic	deviation from WmlComparerSettings (which is wall-clock by default)
DateTimeForRevisions	null → epoch or DateTime.Now	IrDiffSettings.DateTimeForRevisions	explicit value always wins
CaseInsensitive / Culture	false / null	CaseInsensitive / Culture
ConflateBreakingAndNonbreakingSpaces	true	same
WordSeparators	null → default set	WordSeparators
DetectMoves	true	RenderMoves	render-time relabel: the engine always ALIGNS a relocation as a move; this controls whether it is REPORTED as one
MoveSimilarityThreshold	0.8	same
MoveMinimumWordCount	3	MoveMinimumTokenCount
RevisionGranularity	Fine	RevisionGranularity	Fine = engine-native one-revision-per-token-span (byte-stable); WmlComparerCompatible = coalesce/trim/prune to match the legacy comparer's coarser revision set
FormatComparison	ModeledOnly	IrFormatComparison	ModeledOnly reports only modeled-field deltas (false-negative on unmodeled rPr); Full sees every rPr difference. Governs paragraph-property comparison too (block-format-change family — see below)
TrackBlockFormatChanges	true	IrDiffSettings.TrackBlockFormatChanges	public opt-out (cascades to all three slices); detect+track paragraph/table/section property changes (below). Sliced into TrackParagraphFormatChanges/TrackTableFormatChanges/TrackSectionFormatChanges (each defaults equal, so two-way is byte-identical); IrCompositeMerger forces the umbrella OFF but turns all three slices ON, so Consolidate merges every block-format family (B1+B2). Wire: trackBlockFormatChanges (Ops JSON + npm/python)
PreAcceptInputRevisions	false	— (a pre-pass, not an IrDiffSettings field)	when true, runs RevisionProcessor.AcceptRevisions on EACH input before diffing — the first-class "accept-all both sides, then compare" wrapper. See Inputs that already carry tracked changes. .NET-only in v1 (no bridge ripple).
PreserveInputRevisions	false	PreserveInputRevisions	when true, PRESERVE the inputs' own tracked changes Word-style: no pre-accept runs, and the markup renderer carries the RIGHT input's original revision markup through verbatim for content-equal blocks and whole-block inserts (foreign wrappers never re-wrapped; fresh w:ids). Wins over PreAcceptInputRevisions when both are set. One-sided round trip: accept ≡ accept(right) holds, reject ≠ left where foreign markup exists (exactly Word). The DocxCompare selector path sets it.
CompareHeadersFooters	true	CompareHeadersFooters	diff header/footer stories (Word Compare's own default-on granularity — see the Edit script section). false restores the carry-left-verbatim behavior. Wire: compareHeadersFooters (Ops JSON + npm/python types).

Two honest defaults that deviate from WmlComparerSettings

1. Deterministic dates. WmlComparerSettings.DateTimeForRevisions defaults to DateTime.Now — the same compare twice yields different dates. DocxDiff pins a fixed epoch by default so output is reproducible. Opt into wall-clock via Deterministic = false.
2. FormatComparison = ModeledOnly. A w:rPrChange-grade report can only DESCRIBE modeled fields, so a format change driven by an undescribable unmodeled-only rPr flip (w:lang, w:bCs, complex-script toggles) is noise. ModeledOnly collapses that noise; the trade-off is a false negative on a visible-but-unmodeled change (e.g. w:shd run shading). Full restores byte-fidelity comparison.

Paragraph-and-above formatting changes (the block-format-change family)

Word's Compare tracks paragraph-and-above property changes, not just run formatting. DocxDiff produces the native markup Word renders for each (2026-07-03 campaign — strictly ahead of the WmlComparer oracle, which emits none of these):

Scope	Markup	Detection substrate	Round-trip
Paragraph	w:pPrChange (+ w:pPr/w:rPr/w:rPrChange for a changed paragraph mark)	modeled IrParaFormat (incl. direct w:numPr numId/ilvl) under ModeledOnly; the stored fingerprint under Full	inner = OLD (left) pPr, minus w:rPr/w:sectPr/w:pPrChange (CT_PPrBase)
Table cell	w:tcPrChange	IrCell.ShellDigest (whole w:tcPr, folded into cell ContentHash → a shell edit makes the table Modified)	inner = OLD tcPr, minus cellIns/cellDel/cellMerge/tcPrChange (CT_TcPrInner)
Table row	w:trPrChange	IrRow.TrPrShellDigest (trPr-only) for attribution; TrPrDigest (row shell incl tblPrEx) folds into the fingerprint → FormatOnly	inner = OLD trPr, minus ins/del/trPrChange (CT_TrPrBase)
Table row (exceptions)	w:tblPrExChange	IrRow.TrPrExDigest (tblPrEx-only flattened)	inner = OLD tblPrEx (CT_TblPrExBase); a TableRow revision with changed-name "tblPrEx"
Table	w:tblPrChange + w:tblGridChange	IrTable.TblPrDigest / TblGridDigest (fingerprint → FormatOnly)	tblPrChange inner = OLD tblPr; tblGridChange is CT_Markup (bare w:id, no author/date), inner = OLD gridCol run
Section (trailing)	w:sectPrChange on the trailing body w:sectPr	modeled IrSectionFormat (revision) / canonical props-diff excluding references (markup)	inner = OLD section properties (CT_SectPrBase — no header/footer references)
Section (mid-document)	w:sectPrChange inside a paragraph's w:pPr/w:sectPr	IrParagraph.InlineSectionFormat (folded into the paragraph fingerprint + BlockSignature)	inner = OLD inline section properties; a per-paragraph Section revision. One-sided add/remove is structural (untracked); an unmodeled-only inline-sectPr change under ModeledOnly is the same blind spot as run/paragraph formats (right-applied, seen under Full)

Key design points:

• FormatComparison governs paragraphs too. A modeled paragraph delta (jc/indent/spacing/style/numbering) is detected under ModeledOnly; an unmodeled-only pPr delta (e.g. w:shd) is the documented false-negative (untracked right-apply) under ModeledOnly, seen under Full — exactly the run-format trade-off. Table shells and section props are compared canonically (byte-grade) under both policies (there is no modeled/unmodeled split that changes emission there — a documented asymmetry).
• The emitted change carries the FULL old properties, so accept ≡ right and reject ≡ left hold at the property-byte level (canonical, reference/rsid-normalized) for every detected change — the same rule w:rPrChange follows. The reader digests flatten shell children (not the wrapper element), so an empty shell ≡ an absent shell — no spurious change from a render→reject-cycle empty <w:trPr/>.
• Ops carry no new payload. Detection happens at block pairing (an alignment kind flips Unchanged→FormatOnly, or a shell digest flips ContentHash→Modified); the renderers recompute the delta from the source elements by anchor — the established w:rPrChange architecture. The edit-script JSON is unchanged; only the revisions wire gains an additive scope.
• Revision surface. DocxDiffRevision.FormatChange gains Scope (DocxDiffFormatChangeScope: Run default + Paragraph/TableCell/TableRow/Table/Section). Paragraph/section scopes carry modeled property dictionaries; table scopes are digest-grade (ChangedPropertyNames = ["shell"] or ["grid"]). WmlComparerCompatible excludes every non-Run scope by construction (the oracle produces none — keeps the 179-count parity scoreboard meaningful, the hdr/ftr precedent). Full/Fine reports them.
• A consume-side fix rode along: rejecting a w:sectPrChange (or a w:pPrChange on a section-final paragraph) used to drop the section's header/footer references / inline w:sectPr, because those live OUTSIDE the tracked change (CT_SectPrBase / CT_PPrBase). RevisionProcessor now preserves them. See docs/ooxml_corner_cases.md.

Scope coverage (two-way engine — the follow-up A batch, 2026-07-03, closed the gaps below):

• w:tblPrExChange (row-level table property exceptions) is now tracked — IrRow.TrPrExDigest (a flattened tblPrEx-only projection, parallel to TrPrShellDigest) drives a w:tblPrExChange marker + a TableRow-scope revision with the distinct changed-name "tblPrEx"; reject restores the left bytes.
• Mid-document w:sectPrChange (an inline w:sectPr inside a w:pPr) is now tracked — the reader models it as IrParagraph.InlineSectionFormat (folded into the paragraph FormatFingerprint AND IrModeledFormat.BlockSignature so a sectPr-only change classifies FormatOnly under ModeledOnly), and the emit stamps w:sectPrChange inside the paragraph's w:pPr/w:sectPr (not the pPrChange inner — CT_PPrBase excludes sectPr) with a per-paragraph Section revision. A one-sided inline sectPr (added/removed) is a structural change, not a property change — untracked.
• Note-scope and header/footer-scope w:pPrChange already work (they route through the same RenderBlockOp dispatch as the body, with no per-scope gate) — proven by BlockFormatChangeTests (footnote + header pPrChange), not a ceiling.
• TrackBlockFormatChanges is now a public opt-out on DocxDiffSettings (default true; wire key trackBlockFormatChanges, npm/python surfaced).

w:gridSpan / w:vMerge — the chosen scope (Issue #230). Cell grid-span and vertical-merge both live inside w:tcPr, so they are covered by the cell-shell substrate above, not a separate model. The scope decision, closing #230's "property-only table change reads as unchanged" soundness gap:

• Detect (2-way): a gridSpan/vMerge-only change with a stable cell count flips IrCell.ShellDigest → the cell ContentHash → the table classifies Modified, and renders as a native w:tcPrChange with a TableCell FormatChanged revision. accept ≡ right / reject ≡ left hold at the tcPr-byte level (previously invisible — the table read EqualBlock and the edit silently vanished). Pinned by BlockFormatChangeTests.{GridSpanOnly,VMergeOnly}_cell_change_is_tracked_with_native_tcPrChange.
• Compose (Consolidate): a gridSpan/vMerge-only cell-shell edit composes exactly like a cell width/shading edit — sourced from the editing reviewer, consensus/conflict per policy (ComposeCellShell); shell application is not itself w:tcPrChange-tracked in the consolidated output (the documented v1 shell-compose behavior — reject restores text, not shell bytes). Pinned by IrCompositeTableTests.VMerge_only_cell_edit_composes (and the width-shell battery it mirrors).
• Column-structure change (cell COUNT changes — a gridSpan add/remove): detected (the content hash differs, so never silently invisible) and rendered via the table differ's cell alignment. In Consolidate an uncontested column add/remove composes per-cell (w:cellIns/w:cellDel, IrCompositeTableTests.GridSpan_merge_edit_composes); a contested one falls back to a whole-table block conflict. In 2-way the single-toucher path currently lowers a column add/remove to a whole-table del/ins — a pre-existing renderer-granularity limitation (content round-trips), independent of #230's detection soundness.
• Changed-name granularity: a cell w:tcPr change reports ChangedPropertyNames = ["shell"] (the whole w:tcPr is one opaque digest); naming gridSpan/vMerge/width individually would require XML-child diffing the revision path deliberately avoids — by design.

Remaining v1 ceilings (documented + pinned):

• Consolidate MERGES block-format changes (sub-project B, done): reviewers' paragraph (w:pPr, B1), table-shell (w:tcPr/trPr/tblPr/tblGrid/tblPrEx, B2) and section (w:sectPr, B2) format edits compose with per-element attribution + native markup; competing edits conflict per policy. A reviewer who changed BOTH a paragraph's text AND its pPr remains conflict-routed (v1 decision; never a silent drop). See the N-way composite section below.
• Split/merge members do not emit w:pPrChange — a deliberate, principled decline: a split's members are brand-new right paragraphs already tracked by the inserted pilcrow mark (there is no per-member left baseline to diff against), and a merge's non-final members carry deleted marks; a pPr "change" on them is not well-defined and would fight the reject-fuse. Pinned by Split_members_do_not_emit_pPrChange_declined_v1.

Proof: BlockFormatChangeTests, BlockFormatChangeRealDocTests (a real corpus doc mutated across the full table + section family), and the strengthened IrMarkupRendererTests round-trip battery.

Inputs that already carry tracked changes (rule N13 + PreAcceptInputRevisions)

DocxDiff is frequently asked to diff a document that is itself a redline — its body, notes, headers, or comments already carry un-accepted w:ins/w:del/w:moveFrom/w:moveTo/w:rPrChange. The handling is explicit and pinned, not incidental. Characterization tests: Docxodus.Tests/Ir/Diff/RevisionsInInputDefaultTests.cs (default) and PreAcceptInputRevisionsTests.cs (the flag).

The default — diff the ACCEPTED VIEW, carry non-body markup through verbatim.

• Rule N13 (the IR is a revision-free view). IrReader resolves revisions with RevisionView.Accept on a working copy before building the IR (the original bytes are untouched). So the edit script — and therefore GetRevisions/GetEditScriptJson and the body of Compare — is computed over the accepted view of each side: an input's own w:ins/w:del never surface as their own diff, and the produced body carries only THIS diff's revisions, attributed to AuthorForRevisions. At the body level the round-trip already holds against the accepted view (reject(result) ≡ accept-view(left), accept(result) ≡ accept-view(right)).
• The carry-over leak. IrMarkupRenderer assembles the output on a clone of the LEFT package (so styles/numbering/theme/settings/section/media carry over by reuse — what WmlComparer does), and only the body (plus changed footnotes/endnotes and, since the 2026-07-03 campaign, changed header/footer stories) is rebuilt from accept-clean source. Everything else — UNCHANGED header/footer stories, UNCHANGED footnotes/endnotes, styles, the comments part — is passed through verbatim from the original left input. Any pre-existing revision markup there survives into the result, attributed to its ORIGINAL author. This is the documented limitation the inspector's revisionsInInput entry flags.
• Why the leak matters. A leaked pre-existing w:ins in, say, an unchanged header is then rejected by RejectRevisions (it strips the insertion) — so under the default the round-trip does not hold in the carried-over scopes (reject(result) drops header text the accepted view of the left actually contains). Pinned by Default_leak_breaks_the_header_round_trip (whose header is identical on both sides — an unchanged story — so it stays a carry-over even with header diffing on).

The opt-in — PreAcceptInputRevisions (default false). When set, every input is run through RevisionProcessor.AcceptRevisions before it enters the pipeline, so both the IR read and the cloned output package are revision-free on both sides. It is, by construction, exactly Compare(AcceptRevisions(left), AcceptRevisions(right)) — byte-for-byte identical to that wrapper (oracle: Flag_on_is_byte_identical_to_accept_all_then_compare_wrapper). The effect:

• every w:ins/w:del/w:moveFrom in the result is attributable to THIS diff (no stale input revision passed through) in the body, header/footer, note, and style scopes;
• the round-trip holds against the accepted view in those scopes;
• the consumer revision list (GetRevisions) is unchanged — the flag only additionally cleans the rendered package's carried-over parts.

The flag's coverage is exactly what RevisionProcessor.AcceptRevisions processes — the body, headers, footers, footnotes, endnotes, and styles part. Carried-over parts it does not process keep their pre-existing revisions: notably the WordprocessingCommentsPart (a tracked change inside a COMMENT definition survives — pinned by PreAcceptInputRevisions_does_not_flatten_a_revision_inside_a_comment_definition) and the GlossaryDocumentPart (building-blocks / AutoText entries; the renderer clones the whole left package, including the glossary part, verbatim). These are the honest boundaries of the accept-all pre-pass — resolve revisions in those parts separately if they matter. (The narrower NumberingDefinitionsPart carries only style-grade pPrChange/rPrChange, never w:ins/w:del, so it is outside the inspector's tracked-changes scope.)

Applied uniformly to all seven entry points (Compare/GetRevisions/GetEditScriptJson and the four consolidate-family methods, accepting the base and every reviewer). .NET-only in v1 — not yet surfaced on the WASM/npm/python bridges (deferred).

Two honest costs of accept-all (read before enabling). Accept-all is a lossy, opinionated pre-flatten:

1. It flattens pre-existing authorship and change boundaries. Accepting collapses each input's tracked changes into final text, so who made a prior edit and where the prior change boundaries were are lost — the result's authorship reflects only this diff.
2. "Accept all" is itself a policy. It overrides any change a prior reviewer had left unaccepted — including one they had effectively rejected by leaving in tracked form — materializing every insertion and dropping every deletion. To preserve or re-adjudicate the inputs' in-flight revisions, resolve them by your own policy first, then diff. See docs/ooxml_corner_cases.md → "DocxDiff: PreAcceptInputRevisions accept-all flattens prior authorship".

The Word-parity opt-in — PreserveInputRevisions (default false). Word's own Compare does the OPPOSITE of accept-all: pre-existing tracked revisions in the inputs are preserved verbatim in the compare output (original author/date markup intact, verified against Word-oracle outputs), while the text diff is computed over the accepted view. With the flag ON:

• No pre-accept runs (Preserve wins when both flags are set) — the LEFT package's carried-over parts (headers/footers, unchanged notes, styles, comments) keep their markup, which under this policy is the desired behavior, not a leak.
• The renderer reaches back to the ORIGINAL right elements. The renderer's internal IR read accepts the whole working copy first (rule N13), so retained Source.Element provenance is accept-clean; IrMarkupRenderer.BuildPreservedOriginalIndex re-opens the original right package and aligns its body children with the working body via an in-order two-pointer walk that models the document-level accept's only body restructuring (a mark-deleted paragraph merges into the NEXT one; a fully-deleted paragraph vanishes the same way) — a working block maps to a GROUP of originals, and any unmodeled divergence (adjacent-table merges, removed content controls) conservatively stops the walk with a partial map.
• What preservation covers (v1). Content-EQUAL blocks emit the original element(s) verbatim (EmitVerbatim); whole-block INSERTS emit the original with only plain runs wrapped in this diff's w:ins (MarkWholeParagraph leaves foreign w:ins/w:del/move wrappers as-is; MarkParagraphMark/MarkWholeTable keep foreign mark/row markers) — no same-kind wrapper ever nests. Preserved wrappers are normalized (NormalizePreservedClone): fresh w:ids off the render's single counter (range pairs keep one id via a per-render remap) and Word-extension attrs (w16du:dateUtc) dropped. Note scopes ride the same map: footnote/endnote definitions pair by w:id and their child blocks align with the same walk, so equal/inserted note blocks preserve through the same emission paths. NOT preserved (accepted-view render, v1 scope): modified/format-only/split/merge/moved blocks, changed header/footer stories, and the LEFT side's markup in deleted blocks. GetRevisions does not report preserved foreign revisions.
• One-sided round trip, exactly like Word. accept(output) ≡ accept(right) at the text level (foreign dels vanish, foreign ins accepted). reject(output) ≠ left where foreign markup exists — rejecting a foreign w:del restores its text; Word behaves identically, so this is pinned (Equal_paragraph_preserves_foreign_del_and_accept_removes_it) rather than "fixed".
• Wiring. DocxCompare.ToDocxDiffSettings sets it on the engine-selector path (CLI --engine=docxdiff / WASM / npm), so the selector reproduces Word's preservation. Consolidate does not preserve in v1 (the composite renderer builds no map), though Preserve still suppresses the pre-accept there when both flags are set.

Tests: Docxodus.Tests/Ir/Diff/DocxDiffPreserveInputRevisionsTests.cs; corpus proof: page_numbering_examples vs potpourritest (176 foreign revisions in the right input → 103 preserved in the output, zero validator-error delta, accept round trip intact).

N-way composite / Consolidate

The IR engine merges N reviewers' edits — each an independently revised copy of ONE shared base — into a single tracked-changes document, an attributed revision list, a composite edit-script-as-data, and a structured conflict report. This is the IR-native answer to the last WmlComparer capability the engine had not addressed: WmlComparer.Consolidate (the 84 CONSOLIDATE cases in the M2.3 parity inventory, deferred there as "out of v1 scope").

Public surface

Four entry points on public static class DocxDiff (Docxodus/DocxDiff.cs), with the supporting types in Docxodus/DocxDiffConsolidate.cs:

Method	Returns	Purpose
Consolidate(base, reviewers, settings?)	WmlDocument	One multi-author tracked-changes document — each reviewer's edits stamped with that reviewer's own author name (w:ins/w:del/w:moveFrom/w:moveTo/w:rPrChange). The N-way, shared-base counterpart to Compare.
GetConsolidatedRevisions(base, reviewers, settings?)	IReadOnlyList<DocxDiffConsolidatedRevision>	The attributed revision list — DocxDiffRevision's shape plus the contributing reviewer's Author and, on a conflict winner, a ConflictId.
GetConsolidatedEditScriptJson(base, reviewers, settings?)	string	The composite edit script as data: every op additively carries author/sourceReviewer, a conflictId when it won a conflict, and (for a composed paragraph) authoredTokens + sourceRightAnchors; the document gains a top-level conflicts array.
GetConflicts(base, reviewers, settings?)	IReadOnlyList<DocxDiffConflict>	The inspect-before-merge view — the same merge run, surfacing only the conflict list so a caller can review disagreements (and pick a policy) before committing to an output.

Supporting public types (all #nullable enable, XML-doc'd, no static state — author flows per reviewer): DocxDiffReviewer { Document, Author }, DocxDiffConsolidateSettings { Diff, ConflictResolution } (composes — does not inherit — the sealed DocxDiffSettings via Diff), enum ConflictResolution { BaseWins, FirstReviewerWins, StackAll }, DocxDiffConsolidatedRevision (= DocxDiffRevision + ConflictId), DocxDiffConflict { Id, BaseAnchor, TokenStart, TokenEnd, AppliedPolicy, Competitors }, DocxDiffConflictCompetitor { Author, ResultText }. N reviewers, no cap; reviewer LIST ORDER is significant (it determines competitor order and policy tie-breaking). Zero reviewers returns the base unchanged / empty lists. The four surfaces are exposed through every shipping layer — WASM (DocxDiffBridge.Consolidate/GetConflictsJson/GetConsolidatedRevisionsJson/GetConsolidatedEditScriptJson), npm (docxDiffConsolidate/docxDiffGetConflicts/docxDiffGetConsolidatedRevisions/docxDiffGetConsolidatedEditScript), docx-scalpel (docx_diff_consolidate/docx_diff_get_conflicts/docx_diff_get_consolidated_revisions/docx_diff_get_consolidated_edit_script) — all routing through DocxDiffOps, so the wire shapes live in one place.

The merge algorithm

The source of truth is IrCompositeMerger.Merge (Docxodus/Ir/Diff/IrCompositeMerger.cs). The merge builds N pairwise edit scripts — IrEditScriptBuilder.Build(baseIr, reviewer_i) — which all share the base's anchor space AND, within a paired block, the same base token coordinate system. That shared coordinate system is what makes the merge exact rather than heuristic. It then walks the base document in block order; per base block:

• Untouched (no reviewer op, or all EqualBlock) → one base-sourced passthrough.
• One reviewer touched it → that reviewer's op verbatim, authored to that reviewer.
• ≥2 reviewers, all producing the SAME right result → consensus: a single op, authored to the first reviewer. A set of consensus deletes collapses to one delete.
• ≥2 reviewers, all paragraph ModifyBlock token edits with UNCHANGED paragraph properties → token-span composition (ComposeTokenSpans): the per-reviewer token diffs, all expressed over the same base token stream [0, baseTokenCount), compose into one merged authored token-op list. Non-overlapping span edits each land inline under their own author; overlapping spans become a conflict resolved by the policy. The emitted spans tile the base token stream exactly once (a runtime totality invariant via IrCompositeMerger.Invariant — enforced in Release too, not a Debug.Assert that the shipped/CI Release build would strip).
• ≥2 reviewers, all table ModifyBlock edits (row moves uncontested) → per-cell table composition (ComposeTableDiffs): rows align by base row anchor; each reviewer's cell ops pair by base cell anchor (not position, so one reviewer's column change never shifts another's edits), and DISJOINT cross-reviewer cell edits compose inline (each cell authored to its reviewer). A cell edited by ≥2 reviewers RECURSES into the SAME body block/token composition over the cell's paragraph mini-body (MergeBlockStream over the per-reviewer cell BlockOps) — so disjoint words inside one cell paragraph fuse, and same-word edits become a cell-paragraph-anchored conflict resolved by the policy. A reviewer's column ADD/REMOVE composes too: a right-only cell op becomes an authored InsertCell (rendered with native w:tcPr/w:cellIns, removed on reject) and a left-only cell op an authored DeleteCell (w:tcPr/w:cellDel, restored on reject); a delete-vs-edit on the same cell is a recorded cell conflict resolved by the policy. A cell-SHELL (w:tcPr width/gridSpan/vMerge/shading) edit is first-class: the shell digest participates in the cell ContentHash (IrCell.ShellDigest), a changed cell's shell is sourced from its editing reviewer (ComposeCellShell), agreeing shells reach consensus, and competing shells are a recorded conflict. An UNCONTESTED reviewer MovedRow composes (lowered to the del+ins row shape the two-way renderer itself uses). The op's Op.TableDiff is the merged apply/JSON truth; an additive AuthoredRows carries the renderer/revision attribution view. Authored rows/cells tile the base table exactly once (AssertTilesBaseTable, enforced at runtime via IrCompositeMerger.Invariant, not a Release-stripped Debug.Assert). A CONTESTED row move (the same base row also edited/deleted/moved by another reviewer) falls back to a whole-table block conflict.
• Anything else (delete-vs-modify; a reviewer who changed both the paragraph's text AND its w:pPr; a table with a CONTESTED MovedRow; mixed kinds) → a block-level conflict resolved by the policy. (The pPr gate is deliberate: the compose path clones the BASE paragraph's pPr, so a reviewer who also changed pPr would have that change silently dropped — routing such an op to the conflict path preserves and surfaces the reviewer's full edit instead.)

Block-level inserts never conflict. Every reviewer's right-only inserted block appears, slotted immediately after the shared base anchor it follows, ordered by reviewer index — two reviewers both inserting after the same paragraph both appear, attributed. A NATIVE move DESTINATION (see below) is right-positioned content too (null left anchor) and is routed the same way.

Native move composition. Before the by-base grouping, PlanMoves decides, per reviewer move group, NATIVE vs LOWER: a move group (reviewer R, source base anchor S) renders as a native w:moveFrom/w:moveTo iff RenderMoves is on AND R is the only reviewer that touches base block S (so the move does not collide with another reviewer's edit/move on S). Native groups are assigned globally-namespaced move-group ids (one deterministic counter, reviewers in list order then ascending local gid) so two reviewers' independent moves never share a w:name; ApplyMovePlan keeps the native move ops (rewriting the gid) and lowers everything else. The native move SOURCE rides its base anchor through the by-base path (sole-toucher → emitted verbatim, authored to the mover); the native move DESTINATION rides the preceding-anchor (right-positioned) path. COLLIDING moves — move-vs-edit on S, or two reviewers moving the same block — are NOT native: both LOWER to del/ins and record a conflict (see the contested-relocation note below). The markup/revision/JSON renderers already handle native moves with a per-op author override, so no renderer change was needed.

Conflict model + the three policies

A conflict is a base span (a token span, or a whole block) edited DIFFERENTLY by two or more reviewers. The configured ConflictResolution decides what lands in the OUTPUT document; the conflict is always recorded in the data regardless of policy (GetConflicts / the conflicts JSON array / a ConflictId on the winning revision):

Policy	Output at the conflicted span
BaseWins (default)	The base text is kept; every competitor is recorded.
FirstReviewerWins	The first reviewer (list order) is applied inline; the others are recorded.
StackAll	Each competing edit is emitted in reviewer order; all are recorded.

A DocxDiffConflict carries BaseAnchor + the base token span [TokenStart, TokenEnd) (an empty interval = a block-level conflict), the AppliedPolicy, and the per-reviewer Competitors (each with Author and the ResultText that reviewer's edit would have produced — "" for a deletion). Its Id matches the ConflictId on the winning DocxDiffConsolidatedRevision, so conflicts correlate to the revision actually placed in the document.

Multi-author rendering + round-trip

The same renderer backs Compare and Consolidate: IrMarkupRenderer was extended with a per-op author override + per-reviewer source selection (each reviewer's composed INSERT token spans index THAT reviewer's right-token list, so the renderer carries each contributing reviewer's own right paragraph anchor). Single-document Compare output is byte-unchanged by this extension. The round-trip invariant generalizes Compare's accept ≡ right / reject ≡ left: RejectRevisions(Consolidate(...)) content-equals the base (rejecting every reviewer restores the base), and AcceptRevisions(...) content-equals the policy-resolved composite (e.g. base text at conflicted spans under BaseWins, the first reviewer's text under FirstReviewerWins).

Before / after intuition

• Two reviewers editing DIFFERENT sentences of one paragraph. Base: "The cat sat. The dog ran." Alice edits the first sentence, Bob the second. Their two token diffs share the base token stream, so token-span composition fuses them into ONE merged paragraph carrying Alice's edit on sentence 1 and Bob's on sentence 2, each w:ins/w:del attributed to its own author — directly consumable in Word's reviewing pane. (Legacy WmlComparer.Consolidate would instead append two stacked, labeled, colored single-cell boxes after the original — a side-by-side juxtaposition for a human to eyeball, never an inline merge.)
• Two reviewers editing the SAME word. Base: "the quick fox". Alice changes quick→brown, Bob quick→slow. The token spans overlap, so this is a recorded conflict at [BaseAnchor, TokenStart..TokenEnd). Under the default BaseWins, the output keeps quick and GetConflicts returns one DocxDiffConflict with both competitors (Alice→brown, Bob→slow); under FirstReviewerWins the output reads brown (Alice inline) with Bob still recorded; under StackAll both edits emit in order. The conflict data is identical across all three policies — only the document body differs.

Parity outcome — and why the deviation catalog is empty

ConsolidateParityScoreboardTests (Docxodus.Tests/Ir/Diff/ConsolidateParityScoreboardTests.cs) scores all 84 legacy WmlComparer.Consolidate corpus cases (WC001's 10 multi-reviewer rows + WC002's 74 single-reviewer rows): 84/84 reproduce-PASS, 0 deviations, 0 fails, with the genuine-pass floor ratcheted at 84.

The headline finding the scoreboard records: legacy WmlComparer.Consolidate is a juxtaposition/triage tool, not a merge. It keeps the original document intact and, for every changed block, APPENDS that reviewer's labeled revised copy (wrapped in a colored single-cell table under the default ConsolidateWithTable) right after the original — even for a single reviewer. Its accepted body is therefore, per changed block, [revisor label][reviewer's block][original block]: a side-by-side juxtaposition with the revisor name as literal body text. The IR-native engine instead produces a true inline merge. Because the two engines emit categorically different document SHAPES by design, raw accepted-body-text equality is the wrong relation (it never holds — that mismatch IS the supersession). Parity is therefore measured by a sound-semantics metric over normalized body plaintext:

• Single reviewer → accept ≡ that reviewer's document, char-exact (the same accept ≡ right contract Compare obeys; holds for all 74 WC002 rows).
• Multi-reviewer, no conflict → every reviewer's ADDED tokens appear in the accepted body (added-token containment, not whole-body subsequence — composition fuses adjacent edited tokens).

The whole-corpus juxtaposition-vs-inline-merge shape divergence is the deliberate supersession, quantified once in the scoreboard footer rather than catalogued as 84 identical per-row entries. A row is catalogued as a per-row deviation only when GetConflicts reports a TRUE cross-reviewer token-overlap conflict (two reviewers editing the same span differently) — and no legacy-corpus row produces one (every corpus edit is single-reviewer or disjoint-span), so the per-row catalog is empty for this corpus. The conflict-supersession path is instead exercised by the unit suites (DocxDiffOpsConsolidateTests / DocxDiffConsolidateApiTests) and the K-way composite fuzzer (CompositeFuzzTests: round-trip reject ≡ base + the IrCompositeVerifier apply-verifier over 3/4/5-way seeds).

v1 limitations (honest)

• Header/footer scopes are NOT consolidated (2026-07-03 campaign ceiling). The two-way engine diffs header/footer stories (IrEditScript.HeaderFooterOps), but IrCompositeMerger.Merge forces CompareHeadersFooters off for its per-reviewer diffs — explicit and deterministic (nothing is generated, so nothing is silently dropped), pinned by IrHeaderFooterDiffTests.Consolidate_ignores_header_changes_v1_ceiling. A reviewer's header edit is ignored by Consolidate; run a two-way Compare against that reviewer to capture it. Follow-on: MergeHeaderFooterScopes mirroring MergeNoteScopes — simpler than notes (story pairing is by scope/part, no id-map/reference-rewrite machinery).
• Note scopes MERGE across reviewers. The merger builds composite note-scope (footnote/endnote) ops (IrCompositeMerger.MergeNoteScopes): a base-matched note's blocks run the SAME MergeBlockStream dispatch the body uses (disjoint note edits compose, identical ones reach consensus, contested ones are recorded conflicts resolved by the policy; a whole-note delete vs an edit is a delete-vs-modify conflict), and reviewer-INSERTED notes pass through authored. The composite renderer applies the composed ops inside the footnotes/endnotes parts, creates reviewer-inserted definitions under fresh output ids, rewrites reviewer-sourced body references from each reviewer's id space into the base-anchored output space (IrCompositeScript.NoteIdMaps + RenderState.NoteRefClonesBySource; deleted/moved-from content is base-sourced and skipped), then runs the SAME body-order renumber + cross-kind nested-reference sweep the two-way renderer runs. A reviewer-inserted note is all-ins content in that reviewer's id space, so a note-in-note reference INSIDE its definition body (which the body-reference rewrite never visits) is also rewritten to the output id space before the renumber — closing the sub-case where a reviewer inserts a footnote citing an endnote the same reviewer inserts (whose target id becomes a fresh output id, not a base id): without it, that nested cross-kind reference kept the reviewer id and dangled. Same-kind and cross-kind note-in-note nesting therefore stay resolvable on merge/accept/reject across N reviewers (IrCompositeCrossKindNoteTests). Structural ops INSIDE a note (a split/merge/move of note paragraphs) are conservatively lowered to del/ins (content-preserving); native in-note structural composition is a follow-on. Consolidated revisions cover note edits (appended after body ops, footnotes then endnotes), and the parity scoreboard's accept ≡ right metric now includes referenced note texts.
• Multi-reviewer table edits compose per-cell, including column changes and cell shells. DISJOINT cross-reviewer table-cell edits COMPOSE inline (Alice edits cell(0,0), Bob edits cell(1,2) → both land, attributed; disjoint words inside one cell paragraph fuse via the recursion); edits to the SAME cell by ≥2 reviewers become a recorded conflict resolved by the policy. Cell ops pair by BASE cell anchor, so a reviewer's column add/remove composes: an added cell renders with native w:tcPr/w:cellIns (kept on accept, removed on reject), a removed cell with w:tcPr/w:cellDel (removed on accept, restored on reject — Word's own accept semantics absorb the removed cell's grid slot into the preceding cell's gridSpan); a cell delete-vs-edit is a recorded conflict. Cell-shell (w:tcPr) edits are first-class: the shell digest participates in the cell ContentHash (IrCell.ShellDigest), so a width/gridSpan/vMerge/shading-only edit is visible (was: invisible — classified EqualBlock and silently dropped); the composed cell's shell is sourced from its editing reviewer, agreeing shells reach consensus, competing shells are a recorded conflict (BaseWins keeps the base shell). The remaining STOP boundary: a CONTESTED MovedRow (the same base row also edited/deleted/moved by another reviewer) falls back to a whole-table block conflict (no silent loss — the base table is kept under BaseWins and the disagreement is recorded under every policy); an UNCONTESTED row move composes (lowered to the two-way renderer's own del+ins row shape). Notes: shell/tblPr application is not tcPrChange/tblPrChange-tracked (reject restores text, not shell bytes — same class as the two-way renderer's right-shelled table render); a 2-way column ADD in the SINGLE-toucher passthrough still renders via the whole-table fallback (a pre-existing two-way renderer gap, independent of composition).
• Non-colliding reviewer moves/splits/merges render natively; colliding ones are lowered to del/ins as a recorded conflict. A reviewer's MoveBlock/MoveModifyBlock/SplitBlock/MergeBlock whose consumed base block(s) are touched by ONLY that reviewer renders NATIVELY (PlanMoves/ApplyMovePlan for moves with globally-namespaced move-group ids; ApplySplitMergePlan for splits/merges — the same sole-toucher predicate via the shared BuildTouchers map). A native MergeBlock (null left anchor, N consumed base anchors) dispatches through MergeBlockStream's first-consumed-anchor index; the remaining consumed anchors emit no block op but keep insert slotting, and GroupInsertsByPrecedingAnchor advances past the merge's consumed region so a following insert lands after the whole merge markup. A COLLIDING structural op (another reviewer touches a consumed base block) is LOWERED: LowerOneStructuralOp rewrites the move SOURCE → DeleteBlock (retaining MoveGroupId/IsMoveSource as a relocation marker, stripped before emission), the move DEST → InsertBlock, a split → DeleteBlock + N ordered InsertBlocks, a merge → N ordered DeleteBlocks + an InsertBlock — preserving op order; content is fully preserved and round-trips (accept shows the moved/split/merged text, reject ≡ base), and the collision resolves through the existing conflict machinery. Cell mini-bodies run the same split/merge plan (fixing a silent drop: a MergeBlock inside a multi-editor cell reached neither grouping map and vanished). Two reviewers relocating the SAME base block to different places collide on the lowered source-delete → a recorded placement conflict whose PLACEMENT is now resolved by the conflict policy (issue #233): BaseWins keeps the block at its base position (MergeOneBaseBlock emits EqualBlock, not the removal, and every relocating destination is suppressed → accept ≡ base, neither move applied); FirstReviewerWins applies only the first reviewer's (list-order) destination (consensus removal once + the other destinations suppressed); StackAll lands every destination (both/all placements — the legacy behavior). PlanContestedRelocationSuppression selects the surviving destination(s) — each lowered move DEST InsertBlock retains its (reviewer, MoveGroupId) marker so it links back to its source-delete — in lockstep with MergeOneBaseBlock's base-keep-vs-consensus-removal choice (both key off the single IsContestedRelocation predicate). The consensus removal is emitted at most once (never policy-flipped per competitor), so a delete both reviewers made is never duplicated. Under every policy there is no content loss, reject ≡ base holds, and the conflict is recorded for human resolution.
• Reviewers' BLOCK-FORMAT changes MERGE — paragraph (B1) + table-shell + section (B2). The internal TrackBlockFormatChanges is sliced into TrackParagraphFormatChanges/TrackTableFormatChanges/TrackSectionFormatChanges (each defaults equal, so two-way is byte-identical); IrCompositeMerger forces the umbrella OFF but turns all three slices ON. Every family composes by a per-element digest mirroring ComposeCellShell (0 changers → base, all agree → consensus/first reviewer, ≥2 distinct → a recorded conflict resolved by policy; a shell/pPr/section cannot STACK):
  • Paragraph (B1): a pPr-only edit composed by ComposeParagraphFormat over the FULL boundary-normalized BlockSignature (run formats + pPr — NOT the pPr digest alone, which silently drops a competitor's run edit) → native w:pPrChange authored to the winner.
  • Table-shell (B2): per-cell tcPr (existing ComposeCellShell, now with the render stamping the w:tcPrChange marker it previously omitted), per-row trPr+tblPrEx, per-table tblPr+tblGrid — attributed independently by ComposeTableAndRowShells (carried on IrAuthoredRowOp.TrPr/TblPrEx + IrCompositeOp.TableShell, stamped by ApplyComposedShell). Disjoint elements compose; only a contested element conflicts; composes with #250's column-add/remove + row-move. Single-reviewer shells ride the two-way single-source render.
  • Section (B2): the trailing w:sectPr (a document-level element, not a body block op) is composed by ComposeTrailingSection over each reviewer's trailing IrSectionBreak.FormatFingerprint (modeled + unmodeled digest) and stamped as w:sectPrChange (inner = base, hdr/ftr refs preserved) by ApplyComposedTrailingSectPr; an inline (w:pPr/w:sectPr) section change rides B1's paragraph path.
  • reject ≡ base / accept ≡ policy-winner hold at the property-byte level for every family — enforced by the strengthened byte-level verifier (Docs.ShellSection asserted alongside the text projections in CompositeFuzzTests + ConsolidateBlockFormatB2Tests). v1 decision: a reviewer who changed BOTH a paragraph's text AND its pPr is conflict-routed (never a silent format drop); true inline text+format compose is deferred. Pinned by ConsolidateBlockFormatB2Tests.
• Conflict spans are base TOKEN indices (TokenStart/TokenEnd) + BaseAnchor, not character offsets — suitable for machine consumers inspecting the edit script.

Parity status

The engine was developed against WmlComparer as the oracle under a binding method rule: WmlComparer presumed correct per gap; the IR is fixed to match unless an oracle fault is established with concrete evidence. As of M2.6 (this surface):

• GetRevisions parity: 179/179 PASS-or-documented-deviation — 177 genuine count-exact passes plus 2 adjudicated deviations. M2.6 closed the historical catalog entirely (179/179 genuine); the later content-anchored intra-paragraph diff then re-opened exactly two rows, WC-1450 and WC-1470, by making the revision grain deliberately COARSER and Word-aligned (a rewrite is one Del + one Ins region, not per-word). The scoreboard ratchets Pass + Deviation ≥ 179, Pass ≥ 177, and Deviation == 2. The engine optimizes for Microsoft Word's compare output, not WmlComparer count parity, and round-trip is unaffected.
  • WC-1450 in particular (WC023-Table-4-Row-Image, one table row deleted + a "Second " cell edit) is worth knowing because it was mis-filed as an aligner collapse (issue #289). It is not: the body aligns to 5 Unchanged + the table as ONE Modified pair, and IrTableDiffer resolves the table to exactly the authored edit (DeleteRow, ModifyRow$, 2 \times $EqualRow) — pinned by IrSplitMergeTests.WC1450_duplicate_content_body_does_not_collapse_to_delete_plus_insert. The IR reports 2 revisions (the deleted row at its tr: anchor + the "Second " insert), the oracle 7. A whole deleted row is ONE revision at the engine's row grain — which is also what the oracle does everywhere else in the corpus (WC-1140/1150/1460/1660/1670/1750/1760 all pin per-row counts). The oracle's 7 here are not a finer reading of the same edit: it mis-aligns the fixture, reporting content as INSERTED that is unchanged on both sides, plus two null-text revisions.
• Produced-markup parity: floor 39 fixtures round-trip clean (accept ≡ right, reject ≡ left, schema-valid); the round-trip allowlist is 1 fixture (WC-BodyBookmarks endnote→footnote whole-note-store conversion, on which the WmlComparer oracle itself throws — there is no oracle behaviour to match). M2.6 Task 2 closed WC022 (the InOrderRefine same-unid identity reservation); the split/merge markup (below) round-trips on both corpus split fixtures and the synthetic split/merge shapes.
• Round-trip ORDER: accept ≡ right / reject ≡ left hold as exact BLOCK SEQUENCES, not just word multisets — see the in-place order invariant under IrBlockAligner above. Guarded generatively by DocxDiffFuzzRoundTripTests (default 250 seeds, DOCXODUS_FUZZ_SEEDS widens; clean at 2000).
• The only remaining artifact anywhere is that single oracle-crashes allowlist fixture.

1:N paragraph split / N:1 merge (M2.6) — the implemented algorithm

One before-paragraph whose content migrates across N after-paragraphs (the user pressed Enter mid-paragraph), or the reverse merge, is a first-class engine capability. Design: docs/superpowers/specs/2026-06-12-subparagraph-split-merge-design.md (DESIGN-RESOLVED + adversarial review); this section records what the code DOES — the source of truth is IrBlockAligner.DetectOneToManyInGap/FindQualifyingRun/TrimAndGate + IrSplitSegmenter.

Detection (in FillOneGap, after the unambiguous-table-residue rule, before the 1×1-residue rule; gated by IrDiffSettings.DetectSplitMerge, default ON). One side-parameterized worker runs twice per gap — split (singular = left) first, then merge (singular = right); a block consumed by a split group is never reconsidered by the merge scan.

1. Candidates. A gap paragraph on the singular side qualifies iff it is still FREE, or was Modified-paired by THIS gap's SimilarityPair to a plural-side paragraph inside the gap (the pairing is promoted if a window qualifies). Unchanged/FormatOnly/Moved blocks are NEVER candidates: a content-equal pair has zero unmatched tail, so promoting one could only manufacture a false split — this is what preserves the WC022 identity-reservation reject-order invariant (review finding F4.2; regression-tested both directions with detection on).
2. Window enumeration. For each candidate, ascending start × ascending end over maximal CONTIGUOUS runs of eligible plural indices (free paragraphs, or the candidate's own partner), capped at SplitMaxRunLength (8). The first window that passes all gates wins — shortest-first is deliberate: the smallest window clearing the coverage bar absorbs the least foreign content.
3. O(1) length prefilter. Before any scoring, a window is skipped unless its cached content-token total lies in [SplitCoverageThreshold × singularContent, singularContent / (1 − SplitForeignSlack)] — the thresholds' arithmetic implications. Growing a window only adds content, so exceeding the upper bound breaks out of the end-loop. This is what keeps a fully-rewritten G×G gap at G²·O(1) instead of G²·O(LCS) (the adversarial 200×200 fixture's 5-second bound).
4. Scoring (IrSplitSegmenter.Score). One in-order LCS (standard DP, deterministic back-walk tie-break) of the singular paragraph's full token stream against the window's concatenated streams. Coverage = LCS-matched singular CONTENT tokens / singular content tokens; ForeignSlack = unmatched window content tokens / window content tokens (content = non-Separator, non-Textbox; separators participate in the LCS for boundary context but never score).
5. Edge trim (the false-positive guard, review R2). Leading and trailing members with ZERO matched content are dropped, then the trimmed window is re-scored. This excludes an unrelated net-new edge neighbor and edge empty carriers, while keeping INTERIOR net-new members (WC-1830's inserted math paragraph between the two halves) — their foreign content is priced by the slack gate.
6. Fire gates (on the trimmed window). ≥2 members carrying at least one content token; a paired candidate's partner inside the window plus ≥1 other free member; Coverage ≥ SplitCoverageThreshold (0.90); ForeignSlack ≤ SplitForeignSlack (0.34). The thresholds are corpus-swept (IrSplitThresholdSweepTests): the shipped pair sits on a plateau at the grid maximum with ≥1 full grid step of margin on every axis (the F4.1 gate, re-asserted on every run).
7. On fire. Any prior Modified pairing is overwritten; every member's kind/match slots are stamped immediately (no window may reuse a consumed block — the F2.2 overlap ceiling); the group is recorded and its indices leave the leftover lists, so the 1×1 rule and surplus classification see only what remains. EmitEntries collapses each group to ONE alignment entry (IrAlignmentKind.Split at the FIRST member's right position carrying MultiBlocks; Merge at the right block's position), with deletion buckets flushed exactly once per anchored left.

Segmentation (IrSplitSegmenter.ComputeSegmentDiffs, at projection time). The same LCS assigns every singular token to a member: a matched token goes to its partner's member, an unmatched token to the nearest PRECEDING matched token's member (leading unmatched → member 0) — a total, monotone rule. Each slice is re-diffed against its member with the ordinary Myers differ, so every segment diff carries the full token-diff invariant battery over (slice, member) and the partition invariant holds structurally. For a merge the segmenter runs with singular = right, and the builder mirrors each diff (MirrorDiff: Insert↔Delete + span swap) so stored diffs always read left → right.

Surfaces. Apply-verifier: a SplitBlock pushes one reconstructed tuple per member (the existing count/order/ReferenceEquals loop then proves the N rights sit contiguously at the op's position); a MergeBlock pushes one tuple reconstructed from the N members; the cell/note path additionally asserts the produced right-anchor SEQUENCE equals the right block list (the F3.2 strengthening, asserted corpus-wide). Revisions: Fine mode reports each segment's token diff plus one Inserted "\n" per new pilcrow (Deleted "\n" per removed one) — a clean split is visible but claims no content change; compatible mode reproduces the oracle's account — segment 0's inline edits (seam-whitespace-only ins/del suppressed) plus exactly ONE coalesced Inserted ("\n" + Σ(memberText + "\n")) per split — which is what lands WC-1830 at 2 and WC-1450 at 7. Markup: the anchored-split shape — N paragraphs, each member's pPr and segment content, with MarkParagraphMark inserted marks on paragraphs 0..N−2 (deleted marks for a merge, whose non-final paragraphs keep their LEFT pPr); REJECT removes the marks and RevisionProcessor re-fuses the paragraphs, reconstructing LEFT; ACCEPT yields the N right paragraphs. Wire: additive optional arrays only. Fuzzer: SplitParagraph/MergeParagraphs mutation kinds run the own-oracle battery (apply-verify + JSON round-trip + determinism) on every seed; they are excluded from the cross-engine differential class because the engines frame a clean split differently BY CONSTRUCTION (WmlComparer reports the tail as a Deleted+Inserted pair of identical text; the IR keeps it Equal and reports the structural mark — the RelocateParagraph precedent).

Deltas vs the spec worth knowing. (a) The spec's §3.2 "anchored-split" example cell (WC-1450's Second -prefix cell) turned out NOT to be a split at all — its before-paragraph never contained the tail (Score member-match probe: [11, 0]); the oracle's Inserted "Second " + Inserted "When you click…" is the ordinary Modify + InsertBlock account, which the edge trim correctly preserves. (b) The implemented mark placement (inserted marks on 0..N−2, last paragraph keeps the original pilcrow) differs from the spec's §3.2 oracle excerpt but satisfies the same accept ≡ right / reject ≡ left contract the spec adjudicates on (§3.3/§5.5 explicitly allow this). (c) There is no IrSegmentDiff wrapper record (review F1.3) — segmentDiffs is a plain token-diff list. (d) Scope ceilings: N:M and cross-gap splits never fire (one singular side, one gap, by construction); DetectSplitMerge = false restores strict 1:1 op semantics.

Relationship to WmlComparer

	WmlComparer	DocxDiff (IR engine)
Status	Legacy, feature-frozen (kept via explicit selector)	Default (since v8.0.0)
Comparison substrate	Atom streams	Modeled IR (blocks/runs/format records)
Revisions	WmlComparerRevision (OOXML members, no anchors)	DocxDiffRevision (anchor-addressed; no OOXML members)
Move markup	GetRevisions-only post-process; native w:moveFrom/w:moveTo IS produced by the IR markup renderer	native w:moveFrom/w:moveTo
Format change	run-level only (w:rPrChange)	run + paragraph/table/section (w:rPrChange/w:pPrChange/w:tcPrChange/w:trPrChange/w:tblPrChange/w:tblGridChange/w:sectPrChange); modeled-only by default
Diff-as-data	none	edit-script JSON
Determinism	wall-clock dates by default	deterministic by default

Note for readers of wml_comparer_gaps.md: that document's older "native move markup is not generated" / "format change detection is a gap" claims were stale (both shipped in the v6.x line and are produced by DocxDiff's markup renderer here). The gaps doc has been corrected and points here.

Cross-layer ripple

The four-layer ripple (WASM bridge → npm/TypeScript → python host/docx_scalpel) for these three entry points is tracked as M2.5 Task 5 (see the program plan). This document covers the .NET public surface.