EMF+ formatted text in Image Compare

Pinned TortoiseGit: 7338078f8ddd924b8cddee35f512f2286072136d. TortoiseIDiff/MainWindow.cpp:1255 offers WMF/EMF, and Utils/MiscUI/Picture.cpp::Load delegates playback to Windows GDI+. This family implements formatted EmfPlusDrawString in the shared macOS metafile helper used by comparisons, dropped previews and conflict panes.

Record and layout contract

The binary parser follows Microsoft's StringFormat object, variable arrays and DrawString record. Counts are signed where specified, and the complete array extent must match before any layout begins. UTF-16 must contain complete surrogate pairs and XML-compatible characters; malformed input is refused rather than replaced.

CoreText shapes the captured text into actual font glyph outlines. The helper emits vector paths, including fallback run fonts and text matrices, and keeps the original text/font separately in an SVG description. Descriptions are semantic metadata, never a second paint or an external font/resource reference. Text remains vector geometry at arbitrary zoom; existing gradient and texture brushes retain their documented bounded raster paint clipped to that geometry.

The layout supports explicit paragraphs, wrapping/NoWrap, horizontal alignment, line alignment, right-to-left reading order, clockwise vertical text, character tracking, tab stops and first-tab offset, layout clipping/NoClip, whole-line LineLimit, character/word/ellipsis/path trimming, hotkey prefixes and font decorations. Vertical output rotates the completed shaped line once, matching the Microsoft formatting example; vertical-specific CJK glyph alternates remain unverified. Tracking changes shaped advances and positions rather than stretching glyphs. Margins use the recorded em multiplier and the font's converted page-unit size; this interpretation follows actual recorded values and Wine's default font->emSize/6 behavior, rather than the specification's ambiguous inch wording. Tab distances are page units and cumulative: the first is FirstTabOffset plus the first interval, as in the GDI+ GetTabStops example.

Microsoft's managed StringFormatFlags defines bit 0x400 as disabling fallback. Its meaning contradicts the prose in MS-EMFPLUS and parts of the native API page; the implementation uses the managed GDI+ meaning and the requested font's missing-glyph outline when fallback is disabled. The native HotkeyPrefix enumeration controls literal ampersands, hidden markers and visible underlines. Font underline and strikeout use actual font metrics, not guessed line positions. A strikeout font without usable OS/2 metrics is refused.

Digit substitution None (1) is literal. User (0) is implementation-defined in the enumeration; this renderer chooses deterministic unchanged digits for neutral DigitLanguage. Explicit national/traditional substitution and non-neutral DigitLanguage remain refusals. Measured-character ranges are fully bounded metadata and do not alter DrawString paint; this does not add a MeasureCharacterRanges API. DisplayFormatControl remains explicitly unsupported. Glyph-ID DrawDriverString has a separate positioned-text implementation and checkpoint; it is never silently treated as Unicode DrawString.

Beyond the explicit tab list, this native implementation repeats the final interval with its initial phase preserved. Without explicit stops it uses eight actual shaped space advances. These are bounded native continuation/default policies. Microsoft's documentation specifies cumulative explicit stops but not what follows the last one, and Wine's text layout still has a tab-stop FIXME; exact Windows continuation remains unverified. This distinction is retained in the fixtures rather than presented as certified GDI+ behavior.

Non-neutral text Language identifiers are converted from Windows LCIDs to native CoreText language attributes. Unknown mappings fail explicitly; neutral Language uses native font shaping defaults. This is separate from numeric DigitLanguage.

Bounds, ownership and remaining acceptance

Each DrawString is limited to 32,768 UTF-16 units and 256 explicit tab intervals. A conversion has separate 262,144 input-text-unit and shaped-glyph-work budgets, two million outline elements and 32 MiB of serialized text paths. Layout coordinates, margins, tracking, font sizes, line counts and output bounds are validated. Repeated shaping for trimming also charges the work budget. Complex or extreme text can therefore fail explicitly within a parent-owned deadline; those limits are not a claim to accept every otherwise valid GDI+ record.

Zero-size rectangle extents are treated as point-layout anchors for alignment, not as a reason to ignore Center/Far. Oversized blocks retain negative centered offsets rather than silently becoming top-aligned, following Wine gdip_format_string and GdipDrawString. The fitting limit can be unbounded while the alignment extent remains zero. Exact Windows glyph metrics remain separate.

No document path, font file, module loader, network or executable API is exposed to JavaScript. The owning image operation still cancels and drains its exact helper, and rendering preserves original input/conflict bytes. Color/bitmap-only glyphs without vector outlines are refused; Windows-only fonts can differ or substitute through CoreText. Pixel-identical Windows font metrics, all LCID/font variants, vertical CJK typography, physical Image Compare acceptance and signed App Store distribution remain unproven.

The synthetic corpus generator supplies independent records and final-pixel assertions for layout, clockwise vertical orientation, brush paint and clipping, plus malformed framing and resource limits. Batch evidence records actual checks and failed attempts. A focused pass does not certify all metafile drawing or whole-application parity.

The final universal helper and Debug app builds passed. The grouped Core result covers 38 distinct methods across the original group and focused repairs; only the affected six, then two, then one methods were repeated after actual failures or a source-proven alignment correction. Both architectures executed actual baseline/RTL/vertical/point text conversions and malformed UTF-16 refusal. The native consumer family passed with system and packaged Git, including six new formatted-text files and the retained corpus, chooser, previews, conflict and ownership paths. Receivers remained invisible and were reaped; original bytes were preserved. The new source/archive/patch provenance matches the packaged runtime. These scoped results do not establish a full hosted integration pass, all Windows font behavior or signed App Store acceptance of this text subset.