[dsn27]

Part of #412, step 6 (fonts). This PR is stacked on #471 (variable fonts), so its base is feature/variable-fonts. It will be retargeted once #471 is merged.

The comparison

GdiFontComparisonTests (Windows only, #if WINDOWS) compares the ManagedFontProvider with GDI, which is how CADability got its fonts before #412.

Test	What it compares
TheManagedProviderKnowsAllFamiliesOfGdi	Every family name that GDI+ lists must resolve to a managed face, including names that GDI cuts to 31 characters.
FacesMatchGdi	12 common fonts in every real style, covering TrueType, a collection (Cambria), variable fonts (Bahnschrift, Segoe UI Variable) and a STAT-named instance.
AllInstalledFontsMatchGdi	All installed families (131 on the runner), with fewer characters.

FacesMatchGdi checks the metrics, the advance widths, the outlines and the underline and strikeout lines. AllInstalledFontsMatchGdi checks the metrics, the advance widths and a few outlines.

The references and tolerances:

• Metrics: the values of GDI+, exactly.
• Advance widths: GetCharABCWidths, within 1 font unit.
• Underline and strikeout: the last figure of GraphicsPath.AddString, within 2 font units.
• Outlines: GDI's unhinted outlines (GetGlyphOutline with GGO_NATIVE | GGO_UNHINTED), Hausdorff distance within 2 font units.

GDI+ is not the reference for the outlines because AddString returns hinted outlines. In strongly hinted fonts these differ from the design by a few font units, for example 17 units for '2' in Calibri and 42 for 'g' in Constantia. For Segoe Script they are simply wrong, up to 641 units. The managed reader matches the unhinted outlines exactly, with a distance of 0 in every case checked.

What the first Windows run found, and the fixes

The causes were confirmed against the raw font tables with a throwaway diagnostic program on the Windows runner (branch diag/gdi-fonts, not for merging):

1. Family names cut to 31 characters. GDI cuts family names to 31 characters (LF_FACESIZE), e.g. "Bahnschrift SemiBold SemiConden", and texts saved with GDI keep that name. The managed provider now resolves such a name to the only family whose name starts with it.
2. USE_TYPO_METRICS. When bit 7 of fsSelection is set, GDI+ takes the cell (ascent, descent) and the line spacing from the typographic metrics of OS/2, not from usWinAscent and usWinDescent. This affects Bahnschrift, Cascadia, Sitka, Cambria Math, Gabriola, Nirmala Text and Sans Serif Collection. This changes the line layout of texts in these fonts, to what GDI+ did before Get the fonts of texts from a font service of the kernel instead of GDI #468.
3. Underline and strikeout positions. GDI+ centers both lines on the positions of post and OS/2. Before, they were drawn below those positions.
4. The glyph origin. When the xMin of a glyph differs from its left side bearing in hmtx, Windows (and FreeType and fontTools) moves the outline so that the origin is at 0. This affects Times New Roman Italic and Symbol. A composite glyph with USE_MY_METRICS takes the origin of that component instead (Times New Roman Bold 'Ü'). The varied origin of a variable font is taken into account.
5. Names of variable instances from STAT. Windows builds the names from the axis values in the axis ordering of STAT and leaves out the elidable ones. For example, the fvar instance "Light Display" of Segoe UI Variable becomes "Segoe UI Variable Display Light". Without a usable STAT table, the subfamily name of fvar is used as before.
6. MVAR. It is now applied to the metrics of an instance: typographic and Windows ascent and descent, underline, and strikeout. Bahnschrift varies the size of its strikeout line this way. This removes the "MVAR is not applied" limitation of Read variable fonts: gvar, CFF2, avar and HVAR (#412, step 6) #471.

One case is a GDI artifact rather than a reader difference. With FW_NORMAL, GDI emboldens a lighter face such as the instance "Cascadia Code ExtraLight". The test therefore creates the GDI font with FW_DONTCARE for a regular face.

Tests on Linux

Each fix has a test that runs on every platform:

• the origin, by patching the hmtx of Liberation Sans in memory;
• USE_MY_METRICS, through Liberation Sans 'Ü';
• USE_TYPO_METRICS, by patching fsSelection;
• the centered underline;
• the STAT ordering and elision, by patching the STAT of VarTestSans;
• MVAR, against fontTools' instancer values, which were added to VarTestSans.reference.json by make_variable_test_fonts.py.

Mutation checks confirmed that these tests catch the following when they are switched off: the origin shift, USE_MY_METRICS, MVAR, the STAT ordering and the elision.

Results

• Windows (diag/gdi-fonts run of this head): CADability.Tests 675 passed, 2 skipped, 0 failed. This includes the GDI comparison and the reference bitmaps.
• Linux, local: CADability.Tests 651 passed, 12 skipped; CADability.ImportTests 88 passed. CADability.Forms builds.

The Windows CI job (#469) is not part of this stack, so this PR's own CI runs on Linux only. The integration draft #467 runs both.

No public API change. TrueTypeFont and FontVariations are internal.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YXvDW91oXpy7RquGHNxgR5

---

Generated by Claude Code

[dsn27]

The Linux job on 23aa919e failed in CADability.ImportTests, in StepCyclicReferenceTest.import_step_with_representations_that_map_each_other, with a NullReferenceException in ImportStep.Read. The same failure appeared on #468 before.

This failure is not caused by this PR. The parallel STEP import added to the HashSets mappedItems and products from several threads without a lock. The test from #498 imports a file with mapped items 20 times in parallel, and that exposes the race. With the test process limited to 2 cores, the test failed 3 of 30 times without a lock and 0 of 60 times with one.

The fix is #499, against master. I cherry-picked it into the bases of both stacks, #466 and #473, and merged it up through #468, #470, #471, #472 and #479, and through #474 and #480. Once #499 is on master, the change is a no-op in these branches.

---

Generated by Claude Code