fix(docx-compare): emit schema-valid paragraph move markup

[stevenobiajulu]

Problem

The tagged comparison engine emits paragraph move markup that fails the repo's
vendored ECMA-376 Transitional WML schema gate. This is observable in a real
compareDocuments result, not only the lower-level tagged publication helper.

For a two-paragraph reorder with detectMoves: true, the output includes:

<w:moveFromRangeStart w:id="1" w:name="taggedMove1"/>
<w:moveFrom w:id="1" w:author="Range Statistics" w:date="2026-08-23T12:00:00Z">
  <w:p>...</w:p>
</w:moveFrom>
<w:moveFromRangeEnd w:id="1" w:name="taggedMove1"/>

with an analogous moveTo sequence.

scripts/check_emitted_document_schema.mjs reports:

• moveFromRangeStart / moveToRangeStart are missing required w:author and
  w:date attributes;
• moveFromRangeEnd / moveToRangeEnd carry disallowed w:name attributes;
• block-level w:p is emitted inside w:moveFrom / w:moveTo, where the
  vendored schema rejects it.

This shape is not listed in coverage/emitted-schema-known-failures.json.

Acceptance criteria

• Emit paragraph moves using an ECMA-376-conformant structure accepted by the
  repo's vendored Transitional WML schemas.
• Add the appropriate @conformance ECMA-376 edition 5, Part 1 § ... citation
  to the implementation and Allure conformance metadata to the tests after the
  exact governing section is verified.
• Preserve a balanced, ID-consistent move source/destination pair.
• Preserve Accept All and Reject All projections and Word/LibreOffice opening
  behavior.
• Add an end-to-end generated-DOCX schema fixture; do not merely allow-list the
  current invalid output.
• Confirm the fixed markup in Microsoft Word and at least one independent
  consumer oracle.

Premise independently reproduced at 065eef59 on 2026-08-23 with a synthetic
public DOCX and:

node /tmp/revscratch/e2e.mjs
node scripts/check_emitted_document_schema.mjs /tmp/revscratch/e2e

The schema command failed with the errors listed above. No duplicate issue was
found by searching for moveFromRangeStart author date schema or move-schema
terms.

[stevenobiajulu]

Fresh-main classification at 6a7bb7a3 confirms this issue remains narrowly scoped to paragraph-move markup: 1,656 move-range attribute diagnostics plus 2 block-paragraph-in-move-wrapper diagnostics. The remaining public-ILPA schema classes are distinct current defects and are now tracked separately: #944 (6 illegal header/footer relationship references in CT_SectPrChange snapshots) and #945 (16 illegal leaf-level w:del placements). Fixing #941 should not absorb either serializer repair.

[stevenobiajulu]

PR #989 review: residual move characterizations (not closed)

Fresh dynamic Claude Opus 5 integration review at b5f326a73b8e00305a402e3eb263580630afd3f3 approves the repair. Keep these separate from the native/AST mark-consumption fix in #985:

• A selected move endpoint's own paragraph section is consumed with its break, matching the existing insertion/deletion policy. Independent native before/after controls measure sections [10000,12240] -> [12240]; the insertion mirror already does this on main. Source and resolved plain-section controls are among 28/28 validated packages. Add durable move/section controls and independent reader/Word projection evidence before proposing a broader section/binding policy. Word is UNVERIFIED. A separate diagnostic used a malformed headerReference; exclude that diagnostic from package/binding evidence rather than asserting it established schema-valid header behavior.
• Selective author filtering still removes authored range starts but leaves authorless range ends. Before/after main comparison proves this inherited behavior is unchanged; accept-all removes both. A safe follow-up needs pair-aware selection, complete named ranges and independent foreign-author preservation, not a filter that silently resolves unrelated revisions.
• Cross-paragraph bookmark rescue remains an inherited characterization gap, distinct from the repaired locally paired bookmarks around untracked/mixed survivors.
• Native restored moveFrom can retain an empty pPr/rPr husk where AST omits it. This is byte-identical on the pre-repair main baseline, schema-valid, and tracked separately by test(docx-core): characterize terminal paragraph-mark cleanup parity #982's conservative cleanup work; it is not a justification to normalize every unrelated empty paragraph.

Full review and executable probes retained locally at /private/tmp/safe-docx-985-move-mark-20260917/.peer-review/985-integration-988-opus5-20260917.review.txt (probeI1–I8; schema-i1.ts). Public/synthetic only; no document uploads.

No claim that this review establishes general Word, section-layout, selective-group or spanning-bookmark fidelity.

[stevenobiajulu]

Follow-up characterization from the independent Opus review of #985 / PR #989 (reviewed head 3808887f3d5fbabb5b17a9b7240004cfb334b488, main control 24b749524c62f8bf1c9852736482801a129c4b5c): accepting a fully selected deleted paragraph with a bookmark start in that paragraph and its end in the following kept paragraph loses both endpoints. The ordinary del case reproduces on current main, so this is inherited rather than a new move repair regression. Rejecting the analogous ins case preserves both endpoints. Move source/destination probes show the same accept/reject asymmetry, but those initial move probe inputs lacked required range-start dates and are not schema-valid controls; no schema claim is made for them.

Actual executed output for the ordinary controls:

HEAD accept del: texts=[STABLE], bookmarkStart=[], bookmarkEnd=[]
MAIN accept del: texts=[STABLE], bookmarkStart=[], bookmarkEnd=[]
HEAD reject ins: texts=[STABLE], bookmarkStart=[77], bookmarkEnd=[77]

The fully selected paragraph Accept path harvests direct bookmark endpoints without the local-pair guard used by Reject. Preserve this as an open range/bookmark ownership characterization, not a claim that PR #989 repairs spanning bookmarks generally. A wider range/selection policy is outside this focused fix. Local executed probe and output: /private/tmp/safe-docx-985-move-mark-20260917/.peer-review/i995/probe_accept_pair.test.ts and _probe_accept_pair.out. Word projections remain UNVERIFIED.