[cu1ch3n]

This PR compares complete PDFs before cutting review views, so text edits, moved passages and visual changes use one document correspondence. Source ownership determines the relevant edits, and overlapping review regions display shared ink once.

Final validation passed at a2f3b02. The latest Windows and build-progress changes from main are integrated. Adding the engine suite to Windows CI exposed inconsistent repository path identities; the client, server and Git now use the same native canonical paths. The original restart test remains enabled, with an additional directory-alias regression.

Changes

• Preserve reading flow across page headers, floats and physical word breaks. Shared segmentation retains physical glyph footprints and ordinary word boundaries; overlapping extracted rows no longer duplicate glyphs.
• Compare glyph appearance and relative placement using actual TeX box structure. Horizontal box offsets preserve raised/lowered and nested formula changes, while vertical flow allows ordinary paragraph rewrapping.
• Compare non-text paths, images, masks and shading in their shared PDF clip frames. Detect relative drawing movement and text/graphic occlusion, retain margin and cross-column crops, and ignore coherent reflow of independently clipped figures.
• Keep unchanged context in its original colours. Crop only truly blank rows, preserving faint coloured ink and fractional raster coordinates.
• Verify one sufficient edit combination for each optional indirect effect against its final appearance and position. The combination need not be minimal or exhaustive; accept/reject applies only to the selected edit.
• Remove time-based partial comparison results and the sixteen-variant cutoff. Update settings, translations and documentation to match.
• Run the engine suite in Node 22/24 and Windows CI, with real TeX integration and browser checks on Node 22.

Validation

For the final source:

• 457 engine tests passed with no skips on Node 22, including real TeX integration; engine typecheck and production build passed.
• Application tests: 746 passed, 21 platform skips locally; final Node 22 CI passed 745 with 22 platform skips. Application/desktop typechecks, application build and translation checks passed. Twelve appearance combinations and build-progress browser checks passed after integration with main.
• All 18 engine browser tests passed in CI after integration with main.
• Independent reading-flow checks passed 426 combinations, 852 glyph-conservation checks and 450 incorrect-variant controls.
• Actual indirect publishes cover separated scripts and nested fractions/text boxes; negative controls cover ordinary reflow inside minipages and paragraphs.
• The final independent drawing pass passed 12 scenarios, retained all 47,992 changed test pixels in the reported regions, and verified the actual SVG output for faint margin ink.
• All 172 saved-PDF rounds rendered without errors or alignment fallbacks, with zero overlapping strip regions. Synthetic preservation has no new omission flags. The three historical formula coverage flags were independently classified: all actual changed glyphs are retained and correctly marked; every raw pixel outside change crops belongs to unchanged translated glyphs. Original scores remain 27/30.
• The raw manuscript preservation check flags 173 words across nine cases. All 173 were independently classified using source, physical glyphs, raster checks and delivered SVGs: no actual new omission or incorrect annotation was found. Unchanged reflowed text remains in context, which the raw preservation score excludes. The original raw assertion remains failed and is preserved alongside this classification.
• The merged SyncTeX parser produces byte-identical box/baseline arrays across 241 saved builds and 18,389 pages, plus identical results for 20,770 source queries.
• Native Windows engine CI passed 406 tests with 51 platform/TeX skips. Directory-alias recognition and the original detached-server environment restart test both passed; the restart completed in 2.8 seconds. The alias regression fails against the old probe comparison. Final-head CI passed Node 22/24, Windows, all three desktop platforms and the required gate; CodeQL also passed.

The saved-PDF corpus includes 103 manuscript rounds, 39 synthetic rounds and 30 formula cases. Historical raw preservation flags and three formula coverage flags are retained as diagnostics, not relabelled as passing checks. A source-root adapter correction affects four historical build inputs and two rounds; only those two baseline rounds were rerendered with the corrected adapter. Two existing missing non-TeX metadata inputs are recorded without inventing replacements or dropping their units.

Comparison contract

Repeated text can admit multiple valid correspondences. A printed line-end hyphen can be indistinguishable from a discretionary typesetting hyphen. Unscoped drawing primitives share a page frame because PDF does not always retain the author's figure boundaries. Glyph sampling and 2x raster checks have finite resolution; this is not a proof of arbitrary PDF equivalence. Font ink can slightly overhang a highlight rectangle while remaining fully visible inside the crop.

The optional attribution contract is one verified sufficient combination, not all causal explanations. Renderer/compilation failures remain failures and never become successful truncated comparisons. The saved corpus covers direct crops; synthetic actual-publish tests separately cover indirect effects.

All manuscript inputs, screenshots and raw diagnostics remain private and local.