svm: a metafile's text depends on fonts the browser does not have, and combining marks land in the wrong place

[andiwand]

Split out of #772, which is closed. Not a regression — this predates the whole
#779–#793 stack and is unchanged by it.

Symptom

A metafile drawn by StarMath puts a vector arrow over a variable. In the
browser the arrow sits up and to the left of the letter it belongs to.

In odr-private/odt/Vektoranalysis Zusammenfassung.odt, 600 of its 1032
embedded metafiles are affected — 1022 arrows in one document.

What the translator emits

The arrow is U+20D7 COMBINING RIGHT ARROW ABOVE, written as its own <text>
at the position the metafile gives, exactly as the metafile draws it:

<text x="106" y="1085" style="font-family:OpenSymbol;font-size:423;">&#x20d7;</text>
<text x="106" y="1058" style="font-family:Times New Roman;font-size:423;">a</text>

That markup is correct, and renders correctly when OpenSymbol is
present. The arrow only goes astray when it is not.

Cause

A combining mark has zero advance and its outline is authored to the left of
the origin, so that it lands back over the character before it. StarMath
positions the glyph assuming OpenSymbol's own authoring. Every other font
implements U+20D7 as a proper combining mark — which is why it lands
somewhere else.

The font is the whole story. Measured, same markup, one font each:

font	placement
OpenSymbol	correct, over the letter
Arial Unicode MS	correct
serif, sans-serif, Apple Symbols, STIX Two Math, Lucida Grande, Menlo, DejaVu Sans, Symbola, Segoe UI Symbol	all displaced up-and-left

On macOS, LibreOffice's bundled fonts live inside the application bundle and
are not installed system-wide, so a browser never sees OpenSymbol even on a
machine that has LibreOffice.

What does not work

All measured in Chrome against the real OpenSymbol, not reasoned about:

• A generic fallback (font-family:'OpenSymbol',serif) — no effect.
  Browsers do per-glyph fallback for a missing codepoint regardless of the
  stack.
• Naming other symbol fonts — no effect, per the table above.
• font-feature-settings: 'mark' 0, 'mkmk' 0 — no effect, which tells us
  the offset comes from the glyph outline and not from GPOS mark attachment.
  There is no CSS property that reaches it.
• A base character (  before the mark) — worse: the mark combines
  with the space and the arrow disappears.
• Merging the mark into the run of the letter it belongs to — worse: the
  letter's font (Times New Roman) has no U+20D7, so it renders as tofu; and
  forcing the mark to OpenSymbol with a nested <tspan> breaks the shaping
  and puts the arrow beside the letter.

So there is nothing to fix in the markup. The lever is which font the browser
has.

Not vendoring the font

Embedding a subset of OpenSymbol via @font-face would fix it, and the
machinery exists (internal/font, cff/sfnt subsetting, already used for pdf).
We do not want to vendor the font, so this is recorded as the option that
was considered and declined, not as the plan.

What is left to decide

• Document it as a known limitation of svm output, which is honest and cheap:
  the rendering is faithful only where the named fonts are installed, and
  combining marks are the visible case rather than a special one.
• Or let the embedder supply fonts, so an application that ships OpenSymbol
  (the Android and iOS apps could) gets correct output while a bare browser
  does not.
• Or draw the mark ourselves as geometry for the handful of marks that matter,
  which trades a font dependency for a positioning one — we would have to
  decide the offset the font currently decides.

Reproducing

translate "test/data/input/odr-private/odt/Vektoranalysis Zusammenfassung.odt" /tmp/va

then open /tmp/va/document.html in a browser without OpenSymbol installed.
The formulas' vector arrows are displaced; installing OpenSymbol (or rendering
in LibreOffice, whose player is right) places them correctly.