fix(docx): count a table's rules as both editors do, and round a cell and the chip composed in it once

CHANGELOG.md

@@ -8,6 +8,50 @@ follow semantic versioning; release dates are ISO 8601.
 
 ### Public API
 
+- **A DOCX export counts a table's rules as Word and LibreOffice do, and holds a cell's margins and
+  a chip composed in it to the twip the page gives them.**
+  - **A rule takes two and a half twips an eighth of a point, rounded down.** Measured on ten-row
+    tables ruled at three, five, six, seven and nine eighths, both editors stepped each row by its
+    margins, its line and 7, 12, 15, 17 and 22 twips: a rule of 12.5 twips takes 12, one of 17.5
+    takes 17. Counted to the nearest, a row held by its height under a rule of an odd number of
+    eighths stood a twip short of the page's. A rule between two rows is shared in whole twips,
+    the row above it taking half rounded down and the row under it the rest: halved and each half
+    rounded with its margin, a rule of 15 twips came off as 7 and 7, and every row of
+    `NorthlineProposal`'s and `EditorialProposal`'s investment tables stood a twip taller. Tables
+    of one, two, three and six rows ruled at five, six, seven and nine eighths then took the page's
+    height to within a step of Word's grid and to the twip in LibreOffice (measured).
+  - **A cell's top and bottom margins are rounded to the twip together**, the top to the nearest
+    and the bottom to what the two come to less it: rounded apart, two of 32.69 twips came to 66,
+    and with its rules counted so, every row of `CobaltRota`'s staff grid whose day cell held one
+    shift stood a twip taller than the page's.
+  - **A shape composed in a table cell is held, and its label cut, in whole twips rounded down.**
+    The row the cell is in holds the page's height, rounded once; anything in it rounded up grew
+    it: two of `CobaltRota`'s 183.69-twip chips stacked in a day, each held and its label cut at
+    184, stood the row a twip taller. Where a chip's letters keep its label taller than its
+    borders leave room for, Word grows the chip by the rest, and the cell round it now gives that
+    much of its margins back, half above and half below, as far as they hold, where the chip is
+    all the cell holds and the cell all its row holds — as each of `CobaltRota`'s shifts is, in a
+    slot of its own: every staff row with a split shift in an outlined chip stood 3 twips taller.
+
+  `CobaltRota`'s grid now stands where the page sets it, every rule under its top within a quarter
+  point of the page's in Word, where they ran from 0.4pt high to 0.8pt low: its p90 from 0.99pt to
+  0.46pt in Word and from 0.84pt to 0.39pt in LibreOffice on Windows. Its median moves from 0.32pt
+  to 0.38pt (0.23pt to 0.31pt): the shifts' labels stand about 0.3pt low in their chips, where a
+  composed cell's first line is not seated, which the grid standing up to 0.4pt high had hidden.
+  In the DOCX fidelity corpus 10 documents stand nearer the page in Word — `proposal-northline`'s
+  p90 from 0.71pt to 0.43pt, `cv-panel`'s median from 0.43pt to 0.31pt, `invoice-subscription`'s
+  from 0.43pt to 0.33pt — and 10 in LibreOffice on Windows. Seven are further in each, by a twip
+  or a step of Word's grid, where their tables and panels now take the page's height and the
+  lines in and after them show what they had hidden: `invoice-classic`'s notes stand 0.2pt low
+  under its totals table, which had stood 0.13pt high, where the editors leave more space after a
+  table than the page does; `invoice-platform`'s line items, already 0.3pt high in their cells, a
+  twip higher with the rows round them where the page draws them; `letter-panel`'s body, already
+  0.25pt high, a twip higher under its panel; the line items of `invoice-consulting` and its long
+  version, already low, a step of Word's grid lower and a hundredth of a point in LibreOffice;
+  `invoice-metered`'s and `invoice-payments`' a step of the grid, or two to three twips in
+  LibreOffice. `proposal-editorial` in LibreOffice is the document it sets on a new page. The
+  report names nothing more.
+
 - **A DOCX export stands a table's text cells on the page's baselines, and what follows a
   paragraph or a ruled table where the page sets it.**
   - **A row of text cells is moved by its margins.** The page sets a cell's line its face's ascent

docs/architecture/backend-capability-matrix.md

@@ -80,8 +80,8 @@ Payload records live in `core` under
 | Gradient strokes | ✅ `PdfPathPainter` (pattern stroking colour) | ✅ `PptxGradientFill` (native `ln`/`gradFill`) | ❌ |
 | Image — STRETCH / CONTAIN / COVER fit (`ImageFragmentPayload`) | ✅ `PdfImageFragmentRenderHandler` | ✅ `PptxImageFragmentRenderHandler` (COVER via the picture source crop) | ✅ `DocxSemanticBackend.writeImage` (the box comes from `NodeDefinitionSupport.resolveImageDimensions`, the same rule layout applies to `width` / `height` / `scale` and the content-width clamp; CONTAIN is embedded at its fitted size, COVER via the picture source crop as in PPTX, and the picture type is read from the bytes; its left margin and padding are not written, and a picture drawn beside its text is fitted to its box with its padding in it — each named in the report) |
 | Barcode / QR (`BarcodeFragmentPayload`) | ✅ `PdfBarcodeFragmentRenderHandler` (vector: the ZXing bit matrix filled as merged rectangles) | ✅ `PptxBarcodeFragmentRenderHandler` (native freeforms: the same ZXing bit matrix as merged rectangles) | ⚠️ `DocxSemanticBackend.writeBarcode` (a PNG picture of the same ZXing bit matrix through `BarcodeMatrices`, one pixel a cell, in the symbol's two colours with their alpha and at the node's size, its data as the picture's description; it scans, but its data is part of the picture rather than editable, reported `APPROXIMATED`, which also names a link, a transform or its left margin and padding as not carried; an `anchor` is a bookmark on its paragraph; in a page zone it is skipped) |
-| Table rows — resolved cells, row/col spans, two-pass fill/border paint (`TableRowFragmentPayload`) | ✅ `PdfTableRowFragmentRenderHandler` + row grouping in `PdfFixedLayoutBackend` | ✅ `PptxTableRowFragmentRenderHandler` + row grouping in `PptxFixedLayoutBackend` (positioned rectangles, edge lines, and text frames — never native PPTX tables, which re-lay-out content) | ⚠️ `DocxSemanticBackend.writeTable` (a real Word table on the grid `TableGrid` resolves: `colSpan` maps to `w:gridSpan`, `rowSpan` to `w:vMerge`, and the cascaded `DocumentTableStyle` text style reaches the cell's runs; the cell's fill maps to `w:shd` and its stroke to `w:tcBorders` — the engine's default 1pt black rule where the table states none, not Word's thinner grid — its padding to `w:tcMar`, less above and below the room Word makes for the horizontal rules (half of a rule between two rows, the lower row's, to each; the rules above and below the table whole to their row), each as wide as Word draws it, in eighths of a point (a rule between two rows halved to the twip, an odd one a twip short); a row of text cells moved by its top and bottom margins together, from Word's seat four fifths of the way down an exact line onto the page's baselines, by the least its cells' lines call for and no further than the margins hold — not a row holding a composed cell, one the layout did not measure, one in a vertical merge or whose cells call for opposite ways, nor a least call under a tenth of a point — and what the margins leave of a cell's call, half a point and more, moved by its text's position; a row's cells at the row's smallest top and bottom margins, since both editors give every cell the row's largest, the rest of each cell's padding as space above its first paragraph and below its last, down to the largest margin a cell opening with a table or in a vertical merge keeps; the cascaded `textAnchor` maps to `w:vAlign` on every cell and to `w:jc` on a text cell's paragraph, with the engine's default — the vertical middle, on the left, or on the right for a right-to-left cell — and `DEFAULT` at the bottom left, as the renderer draws it; a composed cell is written by the same writers that write its node anywhere, so one built from an image, a list or a table carries it — a nested table is a real `w:tbl` taking the width of the column it sits in, less its own margins and padding, which is the column's rather than the one the page gives it, since the layout reports a composed cell's content under the owner's path; the paragraph Word requires after a nested table is hidden where it ends its cell holding nothing and no space; a translucent fill is flattened against the colour the page paints under the cell, and a translucent rule against the cell's fill where it has one, since `w:shd` and `w:tcBorders` are opaque, and the table's report note names it (`translucency`); Word re-paginates, so the export states where the layout breaks: every row the layout placed is `w:cantSplit`, `repeatHeader(n)` rows are `w:tblHeader` and keep with the row under them, and a row of blocks is kept whole the same way; the paragraph Word requires after a document's closing table is an ordinary one where the last page has room for two lines below it, so a reader can type below the table, and otherwise a point tall with its mark hidden, so it opens no blank page, reported `APPROXIMATED` since text typed at the end then goes into the table's last cell) |
-| Clip region open/close (`ShapeClipBegin/EndPayload`) | ✅ `PdfShapeClipBegin/EndRenderHandler` (CLIP_BOUNDS + CLIP_PATH) | ✅ `PptxClipSafety` + raster fallback in `PptxFixedLayoutBackend` — a provably no-op clip (padded content that cannot be cut) skips the fallback entirely and stays native, editable shapes; a clip that can cut ink renders through the PDF backend into one transparent picture on the clip bounds (pixel-exact, not editable as shapes; run-level link hotspots are not emitted and custom fragment handlers do not apply inside the picture; `Builder.clipRasterFallback(false)` restores unclipped vectors + warning; the raster targets a 2048px long edge, clamped to between native size and 4x, so a region larger than that is rendered at native resolution rather than downscaled — which also means its transient memory grows with the clip instead of stopping at the target (a 3370pt A0-landscape region costs ~45MB while rendering, against ~17MB for anything up to 2048pt); a true vector clip is tracked in [#413](https://github.com/DemchaAV/GraphCompose/issues/413)) | ⚠️ inline fallback + one-time capability warning; the report names a clip that cuts what its layers paint, measured from the layout's fragments by `DocxClipInk` — a clip that cuts nothing (an icon inside its box, a disc's initials, a photo filling its circle) is not named; a picture that fills a container clipped to an ellipse takes the ellipse as its geometry, which both editors crop it to; a badge's glyph — a smaller picture in a painted container that clips it to its outline (`CLIP_PATH`) and holds nothing else but drawing — is drawn by `DocxDrawings` as a floating picture over the outline, where the layout places it, reported `APPROXIMATED` — inside a filled panel the badge and its glyph are drawn in front of the shading; an icon picture beside its text in an unpainted container or a layer stack is drawn the same way; a filled or outlined rectangle or rounded rectangle holding text, composed in a table cell, which has no place in the layout to be drawn at, is written as a panel — a one-cell table in its fill and outline, its corners squared and reported, its row held at least the outline's height less its borders and margins, which keep inside it, and a one-line label the shape centres top to bottom on a line taller than the room Word leaves its content cut alike on both sides to that room, no closer to its letters than three quarters of a point, and seated where the page sets it; the rest of what a composed cell draws (an icon, a tile, a disc) is the table's own drawing and is drawn by `drawCellDrawing` where the layout puts it, anchored as a rectangle is |
+| Table rows — resolved cells, row/col spans, two-pass fill/border paint (`TableRowFragmentPayload`) | ✅ `PdfTableRowFragmentRenderHandler` + row grouping in `PdfFixedLayoutBackend` | ✅ `PptxTableRowFragmentRenderHandler` + row grouping in `PptxFixedLayoutBackend` (positioned rectangles, edge lines, and text frames — never native PPTX tables, which re-lay-out content) | ⚠️ `DocxSemanticBackend.writeTable` (a real Word table on the grid `TableGrid` resolves: `colSpan` maps to `w:gridSpan`, `rowSpan` to `w:vMerge`, and the cascaded `DocumentTableStyle` text style reaches the cell's runs; the cell's fill maps to `w:shd` and its stroke to `w:tcBorders` — the engine's default 1pt black rule where the table states none, not Word's thinner grid — its padding to `w:tcMar`, less above and below the room Word makes for the horizontal rules (half of a rule between two rows, the lower row's, to each; the rules above and below the table whole to their row), each as wide as Word draws it, in eighths of a point, and counted as both editors count it, two and a half twips an eighth rounded down (a rule between two rows shared in whole twips, the row above it taking half rounded down; a cell's top and bottom margins rounded to the twip together); a row of text cells moved by its top and bottom margins together, from Word's seat four fifths of the way down an exact line onto the page's baselines, by the least its cells' lines call for and no further than the margins hold — not a row holding a composed cell, one the layout did not measure, one in a vertical merge or whose cells call for opposite ways, nor a least call under a tenth of a point — and what the margins leave of a cell's call, half a point and more, moved by its text's position; a row's cells at the row's smallest top and bottom margins, since both editors give every cell the row's largest, the rest of each cell's padding as space above its first paragraph and below its last, down to the largest margin a cell opening with a table or in a vertical merge keeps; the cascaded `textAnchor` maps to `w:vAlign` on every cell and to `w:jc` on a text cell's paragraph, with the engine's default — the vertical middle, on the left, or on the right for a right-to-left cell — and `DEFAULT` at the bottom left, as the renderer draws it; a composed cell is written by the same writers that write its node anywhere, so one built from an image, a list or a table carries it — a nested table is a real `w:tbl` taking the width of the column it sits in, less its own margins and padding, which is the column's rather than the one the page gives it, since the layout reports a composed cell's content under the owner's path; the paragraph Word requires after a nested table is hidden where it ends its cell holding nothing and no space; a translucent fill is flattened against the colour the page paints under the cell, and a translucent rule against the cell's fill where it has one, since `w:shd` and `w:tcBorders` are opaque, and the table's report note names it (`translucency`); Word re-paginates, so the export states where the layout breaks: every row the layout placed is `w:cantSplit`, `repeatHeader(n)` rows are `w:tblHeader` and keep with the row under them, and a row of blocks is kept whole the same way; the paragraph Word requires after a document's closing table is an ordinary one where the last page has room for two lines below it, so a reader can type below the table, and otherwise a point tall with its mark hidden, so it opens no blank page, reported `APPROXIMATED` since text typed at the end then goes into the table's last cell) |
+| Clip region open/close (`ShapeClipBegin/EndPayload`) | ✅ `PdfShapeClipBegin/EndRenderHandler` (CLIP_BOUNDS + CLIP_PATH) | ✅ `PptxClipSafety` + raster fallback in `PptxFixedLayoutBackend` — a provably no-op clip (padded content that cannot be cut) skips the fallback entirely and stays native, editable shapes; a clip that can cut ink renders through the PDF backend into one transparent picture on the clip bounds (pixel-exact, not editable as shapes; run-level link hotspots are not emitted and custom fragment handlers do not apply inside the picture; `Builder.clipRasterFallback(false)` restores unclipped vectors + warning; the raster targets a 2048px long edge, clamped to between native size and 4x, so a region larger than that is rendered at native resolution rather than downscaled — which also means its transient memory grows with the clip instead of stopping at the target (a 3370pt A0-landscape region costs ~45MB while rendering, against ~17MB for anything up to 2048pt); a true vector clip is tracked in [#413](https://github.com/DemchaAV/GraphCompose/issues/413)) | ⚠️ inline fallback + one-time capability warning; the report names a clip that cuts what its layers paint, measured from the layout's fragments by `DocxClipInk` — a clip that cuts nothing (an icon inside its box, a disc's initials, a photo filling its circle) is not named; a picture that fills a container clipped to an ellipse takes the ellipse as its geometry, which both editors crop it to; a badge's glyph — a smaller picture in a painted container that clips it to its outline (`CLIP_PATH`) and holds nothing else but drawing — is drawn by `DocxDrawings` as a floating picture over the outline, where the layout places it, reported `APPROXIMATED` — inside a filled panel the badge and its glyph are drawn in front of the shading; an icon picture beside its text in an unpainted container or a layer stack is drawn the same way; a filled or outlined rectangle or rounded rectangle holding text, composed in a table cell, which has no place in the layout to be drawn at, is written as a panel — a one-cell table in its fill and outline, its corners squared and reported, its row held at least the outline's height less its borders and margins, which keep inside it, and a one-line label the shape centres top to bottom on a line taller than the room Word leaves its content cut alike on both sides to that room, no closer to its letters than three quarters of a point, and seated where the page sets it — its row held and its label cut in whole twips rounded down, as the row the cell is in holds the page's height rounded once, and what its letters keep the label past that room taken out of the margins of the cell it is in, where it is all that cell holds and the cell all its row holds; the rest of what a composed cell draws (an icon, a tile, a disc) is the table's own drawing and is drawn by `drawCellDrawing` where the layout puts it, anchored as a rectangle is |
 | Timeline rail — one logical connector line resolved from marker and entry anchors after layout (`ShapeFragmentPayload` per page) | ✅ `PdfShapeFragmentRenderHandler` — one fragment per page, spliced beneath the markers | ✅ `PptxShapeFragmentRenderHandler` — same payload, same per-page fragments | ⚠️ `DocxDrawings` — the rail is read from the resolved layout's pass fragments and drawn per page as a `line` shape, and the markers as the shapes they are, anchored as a rectangle is: beside an entry's text they move with it when the text above is edited |
 | Transform open/close — rotate/scale about fragment centre (`TransformBegin/EndPayload`) | ✅ `PdfTransformBegin/EndRenderHandler` | ✅ `PptxTransformBegin/EndRenderHandler` (group shape; rotation and centre-pivot scaling via the exterior/interior frame ratio) | ⚠️ inline fallback + one-time capability warning; a turned shape container is written upright, and the report names its transform — in its drawn outline's note, or on its own where the outline draws nothing |
 | Anchor markers (`AnchorMarkerPayload`) | ✅ `PdfAnchorMarkerRenderHandler` + `PdfInternalLinkWriter` | ✅ `PptxAnchorMarkerRenderHandler` + `PptxNavigationWriter` (slide-jump hyperlinks resolved after all fragments, so forward references work) | ✅ `DocxSemanticBackend` — an anchor becomes a `w:bookmarkStart` / `w:bookmarkEnd` pair wrapping the paragraph's text, named as Word requires (letters, digits and underscores, starting with a letter, 40 characters); two anchors that clean to one name stay two bookmarks |