fix: render picture-filled text, tab stops, and vertical modes

AGENTS.md

@@ -260,14 +260,16 @@ Oracle case naming:
 
 Case JSON (VBA pipeline): `test/e2e/oracle/cases-full/*.json`. Each declares `kind: “shape”` or `kind: “smartart”` with layout/dimensions.
 
-Python-pptx pipeline cases: `test/e2e/oracle/cases-pypptx/*.json` (202 cases: 60 text,
+Python-pptx pipeline cases: `test/e2e/oracle/cases-pypptx/*.json` (204 cases: 62 text,
 31 shape-adjustment, 28 zero-adjustment flowchart, 2 shape-effect, 1 text-effect, 20 static 3D,
 8 table, 8 formula, 20 composite, and 24 chart).
 Prefix `oracle-pypptx-{category}-{NNNN}-{slug}` where category is `text`, `shape-adj`,
 `flowchart`, `shape-effect`, `text-effect`, `shape3d`, `table`, `formula`, `composite`, or `chart`. Generated by
 `test/e2e/scripts/generate_pypptx_cases.py`; `--case` accepts repeatable exact/glob filters. Text
 IDs 0040-0055 are the CJK wrap/autofit/spacing interaction matrix; IDs 0056-0059 cover
-`defRPr`/`fontRef` text-color precedence.
+`defRPr`/`fontRef` text-color precedence; ID 0060 covers styled soft breaks; ID 0061 covers the
+native tab-stop matrix; ID 0062 covers every non-horizontal DrawingML vertical mode, anchors,
+stacked-character advance, and mixed CJK/Hangul fallback.
 
 ### Step 2: Identify the Shape
 
@@ -682,7 +684,7 @@ Report: `test/e2e/reports/oracle-failures/full-ground-truth-one-shot.json`
 | `scripts/run_all_shapes_eval.py`        | `--shape-id-min N --shape-id-max N` | Batch evaluate shapes via POST `/api/evaluate/{case}`         |
 | `scripts/analyze_edge.py`               | `<case> [--slide N]`                | Canny edge IoU analysis with visual overlay output            |
 | `scripts/one_shot_full_ground_truth.py` | `--macro-host ... --cases-dir ...`  | Bulk generate PPTX+PDF ground truth from oracle cases         |
-| `scripts/generate_pypptx_cases.py`      | `[--case PATTERN]`                  | Generate 181 python-pptx cases across the documented matrices |
+| `scripts/generate_pypptx_cases.py`      | `[--case PATTERN]`                  | Generate 204 python-pptx cases across the documented matrices |
 | `scripts/reflection_metrics.py`         | `--case-report PATH`                | Validate local reflection fields and erasure sensitivity      |
 
 Key mechanics:

CHANGELOG.md

@@ -7,6 +7,11 @@ and this project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.ht
 
 ## [Unreleased]
 
+### Fixed
+
+- Fixed explicit text tabs, all DrawingML vertical text flows, stacked WordArt spacing, CJK/Korean
+  font fallback, and stretched picture-filled runs, including lazy package media.
+
 ## [1.3.0] - 2026-09-14
 
 ### Added

README.md

@@ -570,8 +570,10 @@ geometry are not yet supported.
 
 Text rendering follows the master, layout, placeholder, shape, paragraph, and run cascade. It
 supports theme fonts, CJK text, bullets, hyperlinks, vertical text, superscript/subscript, Office
-percentage spacing, common wrap/overflow combinations, and selected `spAutoFit` growth. Font
-availability remains part of visual-test provenance.
+percentage spacing, explicit left/center/right/decimal tab stops in horizontal left-to-right text,
+and explicit left tab stops in vertical East Asian text,
+solid/gradient/pattern/stretched-picture run fills, common wrap/overflow combinations, and selected
+`spAutoFit` growth. Font availability remains part of visual-test provenance.
 
 ### Tables
 

docs/ARCHITECTURE.md

@@ -228,6 +228,18 @@ host page's `white-space` rule cannot replace the presentation semantics. Explic
 `horzOverflow` and `vertOverflow` values are resolved independently. Paragraph `eaLnBrk="0"`
 uses the browser's unrestricted break opportunity, while the omitted/true default retains East
 Asian typographic rules; both values are explicit so inherited host CSS cannot change the result.
+For horizontal left-to-right paragraphs, explicit `a:tabLst/a:tab@pos` targets are resolved after
+fonts load from the current browser-laid-out cursor. This preserves the text-frame-relative OOXML
+position across paragraph margins, first-line indents, inline or mixed-run tabs, multiple stops,
+and bullet gutters. Left, center, right, and decimal alignments use the measured following field;
+paragraphs without `a:tabLst` retain the browser `tab-size` path. Right-to-left and vertical text
+also retain that fallback, except for the native-verified East Asian vertical left-tab lane, which
+measures the stop on the vertical inline axis.
+
+For `a:bodyPr@vert`, the renderer maps all six non-horizontal DrawingML values to their matching
+column direction and glyph orientation. Stacked WordArt adds PowerPoint's character advance, and
+the CJK sans fallback stack prefers Korean platform fonts before the generic Unicode fallback so
+mixed CJK/Hangul vertical text keeps native-like glyph metrics and column breaks.
 
 The three autofit choices remain mutually exclusive:
 
@@ -248,14 +260,15 @@ whose size comes only from inheritance remain on the bounded fit path. The nativ
 therefore a finite text-box cohort, not a claim of editor-level parity for every PowerPoint autofit
 context.
 
-### Text color precedence
+### Text fill precedence
 
 `TextRenderer` resolves run styles through the existing master/layout/shape/paragraph/run cascade,
-then applies container color options with an explicit local-precedence check. A color declared on
+then applies container color options with an explicit local-precedence check. A fill declared on
 the run wins first; otherwise a `solidFill` on paragraph `defRPr` wins over the shape's resolved
-`fontRef`. The shape `fontRef` is used only when both local levels omit a text fill. The same path
-handles direct `srgbClr` and theme-backed `schemeClr`, so theme lookup remains in `StyleResolver`
-instead of being duplicated by the precedence layer.
+`fontRef`. The shape `fontRef` is used only when both local levels omit a text fill. Solid, gradient,
+pattern, stretched-picture, and no-fill choices remain mutually exclusive. Stretched picture fills
+resolve embedded or lazy media through the current part's relationships and clip the image to the
+run glyphs. Direct `srgbClr` and theme-backed `schemeClr` still resolve in `StyleResolver`.
 
 The native matrix includes positive paragraph defaults, an explicit run override, the no-local-color
 inverse fallback, and square, wide, and tall containers. This boundary verifies color selection; it

docs/TESTING.md

@@ -282,7 +282,7 @@ This generates/reuses ground truth for all SmartArt layouts available on the loc
 
 For text, shape-adjustment, zero-adjustment flowchart, ordinary-shape and text effects, bounded
 static DrawingML 3D, table, composite, and chart interaction cases, use the python-pptx generator.
-It currently defines 202 cases: 60 text, 31 shape-adjustment, 28 flowchart, 2 shape-effect, 1
+It currently defines 204 cases: 62 text, 31 shape-adjustment, 28 flowchart, 2 shape-effect, 1
 text-effect, 20 static 3D, 8 table, 8 formula, 20 composite, and 24 chart cases. Chart cases 23–24
 exercise readable two-dimensional fallbacks for native 3D column and pie charts; they do not claim
 three-dimensional parity.
@@ -368,6 +368,15 @@ Text IDs 0056-0059 cover color precedence independently from layout behavior: pa
 `srgbClr` and `schemeClr` over a conflicting shape `fontRef`, an explicit run color over both, and
 the inverse fallback to `fontRef` when local fills are absent. The cases span square, wide, and tall
 containers and retain exact color values in their tracked OOXML metadata.
+ID 0060 isolates styled soft breaks from the visible run's metrics and bullet color. ID 0061 is an
+eleven-slide native tab-stop matrix covering leading and inline left tabs, multiple stops, bullets,
+mixed-run tabs, the default interval, explicit center/right/decimal alignment, and a vertical East
+Asian left tab. The vertical row uses a horizontal guide on the inline axis and is also guarded by
+browser coordinate assertions for column centering and the 2-inch advance. Its RTL slide remains an
+observation control.
+ID 0062 is a nine-slide native vertical-text matrix covering `eaVert`, `mongolianVert`, `vert`,
+`vert270`, `wordArtVert`, and `wordArtVertRtl`, plus top/center/bottom anchors, 12/24/36 pt stacked
+character advance, and mixed CJK/Hangul fallback behavior.
 
 ```bash
 cd test/e2e
@@ -377,7 +386,7 @@ cd test/e2e
 
 # Generate the focused text matrices. --case is repeatable and accepts exact names or globs.
 .venv/bin/python scripts/generate_pypptx_cases.py \
-  --case 'oracle-pypptx-text-00[45]*'
+  --case 'oracle-pypptx-text-00[456]*'
 
 # Generate only the bounded static DrawingML 3D matrix.
 .venv/bin/python scripts/generate_pypptx_cases.py \

src/renderer/ShapeRenderer.ts

@@ -124,7 +124,7 @@ import {
   resolveThemeFillReference,
   getFocusedGradientStops,
 } from './StyleResolver';
-import { renderTextBody, resolveTextFields } from './TextRenderer';
+import { renderTextBody, resolveTextFields, type DrawingMLVerticalTextMode } from './TextRenderer';
 import { renderCustomGeometry } from '../shapes/customGeometry';
 import {
   getPresetShapePath,
@@ -284,14 +284,19 @@ function expandCssLengthForScale(length: string, scale: number): string {
 function applyVerticalTextFlow(
   el: HTMLElement,
   anchor: string | null | undefined,
-  upright = false,
+  textOrientation?: 'upright' | 'sideways',
   writingMode: 'vertical-rl' | 'vertical-lr' = 'vertical-rl',
 ): void {
   el.style.writingMode = writingMode;
-  el.style.justifyContent = 'center';
-  el.style.alignItems = anchor === 'b' ? 'flex-end' : anchor === 'ctr' ? 'center' : 'flex-start';
-  if (upright) {
-    el.style.textOrientation = 'upright';
+  // In vertical writing the DrawingML anchor maps to the horizontal block axis.
+  // The inline axis still begins at the physical top of the text frame.
+  el.style.justifyContent =
+    anchor === 'b' ? 'flex-end' : anchor === 'ctr' ? 'center' : 'flex-start';
+  el.style.alignItems = 'flex-start';
+  if (textOrientation) {
+    el.style.textOrientation = textOrientation;
+  }
+  if (textOrientation === 'upright') {
     el.style.whiteSpace = 'normal';
   }
 }
@@ -3056,6 +3061,7 @@ export function renderShape(node: ShapeNodeData, ctx: RenderContext): HTMLElemen
       }
 
       let isVerticalText = false;
+      let verticalTextMode: DrawingMLVerticalTextMode | undefined;
       let textAnchor: string | null | undefined;
       const isSingleLineSpAutoFit =
         !!hasSpAutoFit && !hasNormAutofit && isSingleLineTextBody(textBody);
@@ -3127,16 +3133,28 @@ export function renderShape(node: ShapeNodeData, ctx: RenderContext): HTMLElemen
         if (vert === 'eaVert') {
           applyVerticalTextFlow(textContainer, textAnchor);
           isVerticalText = true;
+          verticalTextMode = vert;
+        } else if (vert === 'mongolianVert') {
+          applyVerticalTextFlow(textContainer, textAnchor, undefined, 'vertical-lr');
+          isVerticalText = true;
+          verticalTextMode = vert;
         } else if (vert === 'wordArtVert') {
-          applyVerticalTextFlow(textContainer, textAnchor, true, 'vertical-lr');
+          applyVerticalTextFlow(textContainer, textAnchor, 'upright', 'vertical-lr');
+          isVerticalText = true;
+          verticalTextMode = vert;
+        } else if (vert === 'wordArtVertRtl') {
+          applyVerticalTextFlow(textContainer, textAnchor, 'upright');
           isVerticalText = true;
+          verticalTextMode = vert;
         } else if (vert === 'vert') {
-          applyVerticalTextFlow(textContainer, textAnchor);
+          applyVerticalTextFlow(textContainer, textAnchor, 'sideways');
           isVerticalText = true;
+          verticalTextMode = vert;
         } else if (vert === 'vert270') {
-          applyVerticalTextFlow(textContainer, textAnchor);
+          applyVerticalTextFlow(textContainer, textAnchor, 'sideways');
           appendTransform(textContainer, 'rotate(180deg)');
           isVerticalText = true;
+          verticalTextMode = vert;
         }
 
         if (
@@ -3183,6 +3201,7 @@ export function renderShape(node: ShapeNodeData, ctx: RenderContext): HTMLElemen
             : OFFICE_SINGLE_PARAGRAPH_LINE_HEIGHT,
         ...(fontRefColor ? { fontRefColor } : {}),
         ...(isVerticalText ? { isVerticalText } : {}),
+        ...(verticalTextMode ? { verticalTextMode } : {}),
         ...(hasSpAutoFit && !hasNormAutofit
           ? (() => {
               const hasExplicitSpacing = hasExplicitParagraphSpacing(textBody);