fix(render): compact structural nested-table end paragraphs

[JSv4]

A required unformatted paragraph after a nested table inherits normal line height and space-after, adding a blank line to the enclosing cell and displacing the following body text. Detect that structural paragraph before inherited formatting is expanded, retain its canonical source identity, and render it as a one-twip line with zero paragraph spacing. Explicit paragraph/run formatting, breaks, visible text, body blanks, and nonterminal cell paragraphs retain their lines.

This PR stacks on #1075. That base is independently measured at 93.48448951086479 (12/12 mandatory passes, no regression), at exact commit a388684b2c05100da0e7870eceab6144de42ed8e. Advance only the renderer baseline to that commit so the unchanged zero-tolerance oracle gate protects its improved image cases as well as all other cases. The independent oracle remains pinned at 80df760778b0df293d179a8a2860f343da6df90a; its fixtures, scoring, thresholds, masks, and font policy are unchanged.

Validation: 98 related native converter/inheritance/wrap tests pass. Two authored real-WASM Chromium tests verify compact and intentionally formatted terminal paragraphs, source text order, and a positive PageMap fragment with cell ownership. Four existing picture/column/endnote browser tests also pass. TypeScript checks pass; a fresh WASM build is 5603 KB compressed, under the 5888 KB budget. Authored XML fixtures are independent of the benchmark fixtures.

The unchanged independent oracle measured the cumulative aggregate at 95.13245986806041: starting main 89.73056858212296, picture-wrap base 93.48448951086479, then this table fix. All 12 mandatory cases pass. The zero-tolerance oracle comparison reports No regression in any protected measurement. Nested tables improve from 83.29018819509314 to 99.76989176704953; all other cases retain the measured base results. Calibration scored 100/100. All 15 reference raster hashes, suite digest, metric contract, and measurement/font environment match the original 89.73 run.

Exact tested cumulative commit: 4f5d76d4baf6f06e354c11ee5a1a86378a93a3cb. Successful benchmark run and complete evidence artifact include all result/check JSON, shared references, PDFs, reference/candidate rasters, overlays, source-bound build/runtime provenance, and comparison findings. Artifact archive SHA-256: d50e381087e512bbcb6eec579d113b2a868d47b3d5ace2c114b4d5edb8fb4351.

Case	Starting main	Picture-wrap base	Cumulative candidate
columns-continuous	95.0258	95.0258	95.0258
columns-balanced	95.9204	95.9204	95.9204
columns-break	96.4066	96.4066	96.4066
paragraph-contract	88.0953	88.0953	88.0953
paragraph-heading	98.9772	98.9772	98.9772
paragraph-keep	83.7345	83.7345	83.7345
notes-footnote	95.6630	95.6630	95.6630
notes-endnote	98.8841	98.8841	98.8841
tables-nested	83.2902	83.2902	99.7699
tables-margins	99.7420	99.7420	99.7420
images-square	91.2869	93.5285	93.5285
images-tight	56.3331	91.6307	91.6307

Reproduce by running .github/workflows/rendering-fidelity.yml at this head, or use the workflow's pinned provisioning/build/reference/measurement commands. Download the evidence with gh run download 38186629726 --name independent-rendering-fidelity. The environment remains LibreOffice 25.8.7.3, Node 22.23.3, Playwright 1.57.0 / sandboxed Chromium 143.0.7499.4, and approved Liberation fonts. The original main baseline is 0289a89ddadff89070563652fcf95897e7656b1b.

The goal above 98 remains unfinished. Paragraph line placement/justification and two-sided picture contour flow still require work. Native, npm, platform builds, performance, and the independent benchmark pass; the full Playwright CI also passes. The Claude review action failed before producing review comments with is_error:true and no model usage.