fix(diff): keep tables anchored when a paragraph move is ambiguous (#229)

[JSv4]

Closes #229.

Problem

When a reviewer moves a paragraph across a table boundary and a second reviewer edits a disjoint cell of that table, DocxDiff.Consolidate collapsed the whole table to a block-level conflict instead of composing the two edits independently.

Root cause is in the pairwise aligner, not the composite merger. A reorder like [A, table, B] → [A, B, table] is genuinely ambiguous:

• "paragraph B moved up past the table", or
• "the table moved down past B"

Both cost exactly one relocation and round-trip identically in a two-way Compare. IrBlockAligner's LIS spine has two equal-length solutions here and the patience-sort tie-break arbitrarily picked the one that relocates the table. That spurious table-move then collides — in the N-way merge — with the second reviewer's disjoint cell edit, and IrCompositeMerger falls back to a whole-table block conflict (BaseWins keeps the base table; the cell edit is recorded in the conflict rather than composed per-cell). No data loss, and reject ≡ base always held — it was a precision/quality limitation.

Proof it existed

IrCompositeTableMoveBoundaryTests reproduces the issue's exact sketch (base paragraph A + 2×2 table + paragraph B; Alice moves B before the table; Bob edits cell(1,1)). Before the fix the test failed on all three conflict policies with a tbl:…-anchored DocxDiffConflict (Alice's empty competitor + Bob's whole-table d BOB competitor) — matching the issue verbatim.

Fix

A structural-anchor tie-break in IrBlockAligner: among equal-length spines, keep the most non-paragraph blocks (tables / section breaks / opaque blocks) anchored and relocate the lighter paragraph instead.

• Implemented as a maximum-weight longest-increasing-subsequence where each candidate weighs BIG + (structural ? 1 : 0) with BIG = candidates.Count + 1. Because BIG dominates any achievable bonus total, cardinality stays the primary key — the number of relocated blocks is unchanged; the bonus is a pure tie-break.
• Runs O(k log k) via a Fenwick prefix-max and is fully deterministic.
• Guarded so it fires only when the plain patience-sort LIS relocated a structural block — the paragraph-only common path is byte-identical (zero churn).

This fixes the root cause at the aligner (also yielding cleaner two-way markup — a paragraph moves, not a whole table) rather than patching the composite merger.

Result

The paragraph move and the disjoint cell edit now compose independently: conflicts == 0, a native/lowered paragraph move plus per-cell table compose, reject ≡ base, no content loss — the issue's acceptance criteria.

Testing

• New IrCompositeTableMoveBoundaryTests (all three ConflictResolution policies) — red before, green after.
• Base two-way parity holds at 179/179 (IrParityScoreboardTests).
• Full suite green: 2606 passed / 0 failed / 3 skipped.
• Release build (TreatWarningsAsErrors) and WASM_BUILD compile clean; no warnings from the changed file.

Files

• Docxodus/Ir/Diff/IrBlockAligner.cs — the tie-break (guard + weighted LIS).
• Docxodus.Tests/Ir/Diff/IrCompositeTableMoveBoundaryTests.cs — regression test.
• CHANGELOG.md, docs/architecture/ir_diff_engine.md — documentation.

No public API change, so no WASM/npm/python wrapper ripple was needed.

🤖 Generated with Claude Code

---

Generated by Claude Code