fix: preserve declared chart series stroke geometry

assets/bugfixes/issue-1578/README.md

@@ -24,9 +24,10 @@ In the red control's page-2 trace, native paints the axis before the marker.
 The old output reversed that order; the fixed output agrees with native.
 All five before/after cases preserve every paint primitive, including its
 geometry and paint, and have identical page-1 traces. Layout and text-layer
-reports retain their baseline findings. Non-whitespace text matches native;
-the stricter text-layer report retains eight extra space-census entries over
-the five-page collection, so it is not an exact text-layer match.
+reports retain their baseline findings. Non-whitespace text matches native with
+`pdftotext -layout`. Default extraction retains table reading-order differences
+and eight extra space-census entries over the five-page collection, so it is
+not an exact text-layer match.
 
 A portable compiled-frame regression covers opaque, hollow and disabled
 markers at two overlapping values. The four enabled-marker cases fail before

assets/bugfixes/issue-1590/README.md

@@ -0,0 +1,84 @@
+# Declared chart-series stroke geometry
+
+Chart-series lines and their legend samples now retain the series' declared
+round, flat, or square cap and round, miter, or bevel join. Miter limits are
+carried through to the rendered stroke. Values, color, width, marker styling,
+and all other paint primitives remain unchanged.
+
+## Native evidence
+
+Source: `tests/fixtures/xlsx/issue_1181_fit_to_height.xlsx`, SHA-256
+`2b4a2d8dceda58758593c88409875efbda05780559154c02bd13fef4f7a1c65b`.
+Native Excel 16.112.3 and the converter use Excel's DFonts. The cash-flow
+series declares 2.25pt, fitted to 1.755pt on the worksheet. Before is main
+`38a69278f2ab911eaa258845d0191cc65fe13bd6`; its complete paint traces match
+the frozen #1578 candidate used to prepare these probes.
+
+The original and unchanged re-zip control have identical complete native and
+converter paint traces. One-factor controls change the local cap to flat or
+square, or the join to miter or bevel. A separate omitted-limit miter control
+has complete native and converter page traces identical to the explicit
+`lim="800000"` control: native exports limit 8. This is measured Excel behavior.
+All seven required cases agree with native after the fix. Every native
+workbook exports two sheets and passes the integrity gate. The exporter
+returned -128 only at its final quit; completed sheet PDFs were preserved,
+united in workbook order, and independently validated. That exporter bug is
+tracked in #1592.
+
+Fifteen before/after cases, including related-style controls, preserve every
+paint primitive except the cash-flow stroke's cap, join and miter limit. Their
+first pages are wholly unchanged. Layout and text-layer findings are identical
+to the baseline. Non-whitespace text matches with `pdftotext -layout`. Default extraction
+retains baseline table reading-order differences and ten extra space-census
+entries, so this is not an exact text-layer match.
+
+A portable compiled-frame regression covers six cap/join/limit combinations
+across line, mixed column/line, and radar charts. It reads both plot and legend
+strokes after compilation, verifies their widths, and prevents nested marker
+or data-point styles from replacing the series declaration. The regression
+fails before the implementation and passes afterward.
+
+## Collection
+
+| Page | Source |
+| --- | --- |
+| 1 | Original workbook, first printed sheet |
+| 2 | Original workbook, second printed sheet; round cap and join |
+| 3 | Flat-cap control, second printed sheet |
+| 4 | Square-cap control, second printed sheet |
+| 5 | Miter-join control, second printed sheet; limit 8 |
+| 6 | Bevel-join control, second printed sheet |
+
+Page 5 also represents the omitted-limit control, whose complete paint traces
+match the explicit-limit control. Full-page evidence uses pdftoppm at 300 DPI,
+progressive JPEG quality 86, metadata stripped before restoring the density,
+and a 0.5pt fine-detail layout threshold. All 2,533 material clusters have explicit passing dispositions across the
+six accompanying reports. Full contexts, original-scale endpoint/join/marker
+crops and final JPEG typography/rule crops were visually inspected.
+
+Remaining independent findings: chart frame #1272; fitted drawing origin
+#1542; column width #1543; fixed-row text #1545; title-row height #1550;
+wrapped-note pitch #1551; footer baseline #1552; worksheet rules #1564;
+missing chart separators #1566; cached marker value #1577. Native marker and
+sparkline bitmap edges also differ from vector output. Omitted local stroke
+properties do not yet inherit the related chart style (#1593); separate
+style-only round/flat/square controls reproduce that gap. This change preserves
+explicit local properties and does not invent a universal round default.
+
+## Reproduction
+
+Run `python3 scripts/probe_harness.py scripts/probes/issue-1590-series-stroke.json --backend office`.
+The portable spec recreates all five measured variant packages byte for byte.
+Check each native export with
+`python3 scripts/check_gt_integrity.py NATIVE.pdf --source SOURCE.xlsx --json`.
+Build each version with `cargo build --locked -p office2pdf-cli`, preserve its
+binary, and convert with `office2pdf SOURCE.xlsx -o OUTPUT.pdf --font-path FONT_DIRECTORY`.
+Use Microsoft Excel.app's `Contents/Resources/DFonts` directory. Assemble
+collections in the table's order with `mutool merge`.
+
+Run `python3 scripts/compare_layout.py GT.pdf AFTER.pdf --json --audit --fine-shift 0.5`
+and `python3 scripts/compare_text_layer.py GT.pdf AFTER.pdf --json`. Their existing
+findings are described above. For every page run
+`python3 scripts/compare_render.py GT.pdf AFTER.pdf --page N --dpi 300 --fine-shift 0.5 --artifacts-dir DIRECTORY --cluster-report PATH`.
+Inspect full pages, matched crops and pixel sweeps, then rerun with
+`--cluster-dispositions PATH --strict-clusters` using reviewed exact IDs.