View Markdown source ↗

EMF+ painting interoperability

This is the painting extension to the metafile decoder, with TortoiseGit pinned at 7338078f8ddd924b8cddee35f512f2286072136d. It addresses reusable image attributes, image placement encodings and gradient brushes within the existing IDiff image panes. It introduces no new image dialog or repository conversion. Original bytes remain authoritative.

Implementation and validation are separate: the first Core checkpoint exposed painting failures, followed by focused renderer diagnostics and production repairs. The final source-pinned runtime and Debug app built successfully; all 18 metafile Core methods and the native consumer group with both Git engines passed. Grouped QA evidence retains all failed attempts and the exact scoped results. Earlier metafile evidence predates this extension and cannot establish these painting results.

Independent writer comparison

The user-provided layout2vector was inspected at 2e0a96b9f87f2a8c0ff649963a67dfccd73bf5bb under MIT. It exports EMF and EMF+; it is neither TurtleGit's decoder nor a WMF implementation. No generated browser/Paint output from that project is adopted as an original TurtleGit fixture or proof of Windows acceptance.

Its EMF+ writer provides concrete interoperability cases:

The classic EMF writer and its scripts/generate-paint-emf-probes.mjs additionally distinguish STRETCHDIBITS plus world transforms, clipping, text and compound path fills. The probe README requests manual Paint observations; its existence alone is not a successful Paint result.

Format contracts

Microsoft's ImageAttributes object specification defines wrapping, clamp color and rectangle/bitmap clamping. Supporting the writer's default object does not imply color matrices, gamma adjustment or every image effect. Reserved fields must not become invented drawing options.

DrawImage uses either integer or floating-point destination rectangles. DrawImagePoints adds compressed/relative point encodings and maps its three corners to a parallelogram. Relative points are not interchangeable with fixed-width signed 16-bit coordinates. Source units remain pixels. Record flags must be interpreted before selecting the point layout; truncation must fail rather than shift the remaining fields into plausible coordinates.

Linear gradient data and path gradient data have different boundary and optional-data layouts. The BrushData flags select transforms, preset colors, blend factors and focus/gamma behavior. Path-gradient preset/factor positions follow the GDI+ playback API: SetInterpolationColors and SetBlend run from boundary position 0 to center position 1. SVG radial stops run the other way, requiring reversal. The generic wire-structure description alone is not used to invert actual GDI+ playback behavior. This also means accepting a writer's records does not certify that writer's intended CSS appearance.

Image attributes and placement subset

The batch implements referenced type-8 ImageAttributes with RectClamp (0). Wrap values 0–3 emit explicit visible image copies with the correct negative and reflected tile phase; native SVG pattern tiling is not used. Inverse current world-transform bounds limit copies to the actual viewport, with caps of 16,384 tiles per record and 65,536 per conversion. Before concatenation, tile markup reserves count * (escapedPNGhref.length + 4096) bytes against an aggregate 64 MiB budget; count limits alone would allow a small bitmap to expand into multi-gigabyte inline SVG. Clamp (4) paints the requested ARGB color only outside the source image extent before placing the original image. Keeping the clamp area separate avoids painting an opaque backing under translucent source pixels. The Clamp BOOL (0 or 1) is ignored under the documented GDI+ 1.0 contract. The absent-attributes sentinel is 0xffffffff; an explicit table index must be 0–63 and name a present ImageAttributes object, rather than an unrelated object type.

Both image records validate pixel source units, bounded nonzero source extents, record lengths and referenced images. DrawImage supports floating-point and signed 16-bit destination rectangles. DrawImagePoints supports floating-point, signed 16-bit and relative signed 7/15-bit point encodings before applying the current graphics transform and source crop. The relative flag is 0x0800; 0x2000 denotes an image-effect route and is refused. Relative image placement does not imply support for the separate relative path encoding. Source-copy composition is implemented by the separate SourceCopy family; image effects remain refused. Generic embedded metafile images, textures and applicable attributes now follow the embedded-image family. Up to three wire-alignment bytes are ignored, without treating additional bytes as an extended attributes structure.

Gradient subset

The local turtleGitPaint.js renderer module adds linear and a restricted path brush through existing shape/text fill resolution. Linear gradients accept endpoint colors, preset color stops or horizontal blend factors, plus independent nonsingular affine brush transforms. The gradient axis is horizontal in brush coordinates; the brush matrix and current graphics transform compose separately. Wrap values 0/2 repeat, 1/3 reflect, and 4 pad. Vertical factor arrays are bounded and consumed but do not change painting, matching Microsoft's documented Windows GDI+ behavior. The supported subset requires both reserved endpoint-color fields to duplicate their corresponding endpoint colors. Microsoft's general structure text says these fields should be ignored, but its Windows product notes document empty rendering when both duplicates mismatch. The stricter rejection of a single mismatch is a conservative unsupported-case policy; it is not a claim that Windows necessarily paints nothing for a single mismatch.

Path gradients are accepted only for a centered, affine image of the canonical four-cubic ellipse: a 13-point Float32 boundary with the expected point types, a matching bounded-tolerance center, and uniform boundary ARGB colors. Float32 coordinate precision and canonical-shape tolerance must each fit within 100 parts per million of the conservative minimum ellipse axis (abs(det(U,V)) / hypot(Ux,Uy,Vx,Vy)); otherwise the brush is refused. The bound also rejects excessively ill-conditioned axis ratios. This prevents a huge coordinate offset from admitting visibly noncanonical or off-center geometry after a later world translation/scale. They support Clamp wrapping, optional affine brush transform, preset colors and horizontal blend factors. Positions/factors retain boundary-to-center semantics; the paint evaluator reverses their sampling orientation. Painting beyond the boundary remains transparent rather than extending the boundary color across the filled shape.

Stop arrays contain 2–4096 ordered finite positions with endpoints 0 and 1; duplicate positions remain valid hard transitions. Invalid factors/counts, conflicting preset/factor flags, singular matrices, unsupported versions and unexpected object bytes fail. General paths/point boundaries, off-center radial brushes, poorly conditioned huge-offset ellipse coordinates, nonuniform surround colors, focus scales, gamma correction and radial tiling remain unsupported. This accepted ellipse subset covers a concrete writer encoding without approximating arbitrary path-gradient geometry.

Native display representation and failure repairs

Focused native raster diagnostics demonstrated that CoreSVG ignored gradient repeat/reflect, shifted gradient colors on PNG round-trip, and tiled bitmap SVG patterns incorrectly. Those observations required production painting changes; they were not resolved by removing pixel assertions. A separate native SVG alpha issue interpreted rgba(...,1) as byte alpha 1/255.

Gradient objects create semantic SVG gradient definitions in the live drawing context, but actual brush color is evaluated analytically into a bounded PNG through the existing native __turtleBitmap adapter. SVG snapshots, including the final output, serialize only reachable definitions and their transitive references. Unreferenced semantic gradient definitions are therefore omitted; their absence does not indicate a solid-color replacement. The live definitions remain available for later object selections and SaveDC/RestoreDC. The original fill geometry remains a vector clip. Gradient text uses white text in a luminance mask because native SVG clip text was not honored. Conservative path-control-hull/rectangle/ellipse-arc bounds receive one pixel of padding; text falls back to the viewport bounds. Brush/current-matrix/crop caching avoids repeating an identical paint raster.

Each paint raster is limited to 16,777,216 pixels and 64 MiB of raw color data; generated gradient rasters together are limited to 67,108,864 pixels per conversion. Wrapped linear gradients use twofold sampling when the per-raster budget allows it to reduce native upsampling seams. Global alpha is applied to the placed image. This is an intrinsic-output-resolution display raster, not a resolution-independent vector gradient or a claim of exact rendering at unlimited zoom. Reopening/decoding original metafile bytes remains separate from preview zoom. Solid geometry/text and clipping retain their existing vector route.

The final SVG color pass emits RGB color and explicit opacity instead of relying on CoreSVG's ambiguous RGBA alpha handling. Clamp color likewise uses separate opacity. This preserves opaque and translucent source fills/strokes without changing captured metafile bytes. Unsupported operations continue to fail rather than presenting a successful incomplete picture.

Authored verification scope

The original MIT synthetic generator has produced seventeen new EMF files for the following cases. The generator was rerun for focused fixture corrections after the failed checkpoint; all eighteen focused Core methods subsequently passed. A subsequent native clamp-color assertion required the pixel-inspection helper to use an explicit sRGB byte buffer: NSBitmapImageRep.colorAt did not preserve the PNG profile in that comparison. Color thresholds were retained. This fixture-helper correction is separate from the production rendering repairs. These fixtures contain original data rather than copied writer output, images, fonts or SVGs. Eighteen focused Core methods and two added native consumer samples are the final focused verification scope:

Scope and remaining work

The subsets above passed this batch's source reconstruction, focused Core pixel checks and native consumer checks. These results establish only the recorded subset and workflows; they do not certify complete Windows graphics parity.

The later region and pattern extension addresses region painting and hatch/bitmap texture brushes with separate evidence. Later formatted text, indexed/16-bit bitmap and pen stroke and custom-cap extensions have separate evidence. Explicit gaps remain for driver/glyph strings, other raw bitmap formats, uncommon classic raster operations and Windows font/antialiasing identity. Any newly accepted gradient subset must retain explicit refusals for its remaining optional fields, malformed boundaries and unsupported painting. Three older image fixtures now use the explicit 0xffffffff absent-attributes sentinel; this preserves their original crop/text/bitmap purpose while separate fixtures exercise real attributes. Native consumers add transformed gradients and default attributes to the retained WMF/classic EMF/EMF+ samples.

All helper input/output limits, cancellation, timeouts and native ownership rules continue to apply; no painting feature grants filesystem or network access.

The compressed-path continuation now shares the wire reader for that same accepted ellipse boundary. Short/relative encoding does not relax the thirteen-point geometry, center or tolerance rules above.