fix: load a text style's font before writing it, and let its typography bind to variables

[awdr74100]

The last shape #164 exposed on read but could not write back — and, found while measuring it, a
bug that had nothing to do with bindings.

The bug found on the way

update_text_style with nothing but a fontSize fails:

in set_fontSize: Cannot write to node with unloaded font "Inter Regular".

So does lineHeight, and so does letterSpacing. Reading a style does not load its font, and
these handlers only loaded one when the call also passed fontName — so updating a size on any
style the caller had not also re-fonted in the same session simply failed. That is most styles.

create_text_style had the same hole with a worse ending: it wrote after creating, so the throw
left an orphan style behind in the design-system panel — reproduced live, style intact in the file
with the default fontSize: 12 it never got past.

Both fixed by loading the style's own face before any typography write. That load is hoisted to
just after creation, the earliest possible when fontName is omitted (the face is then whatever
default Figma put on the fresh style). Safe where the caller-input load is not: a default face
cannot fail on a bad argument.

The test that asserted the opposite went with it — it read "a size-only change touches no glyphs
→ no font work", which is a plausible thing to believe and not what Figma does.

Text-style bindings

create_text_style / update_text_style now take the same boundVariables get_styles reports.
A text style's values are scalars, so unlike a paint or an effect there is no per-object level to
hang them on — the style itself carries them.

Two differences from the paints and effects in #167, both because a text style write is a patch
rather than a whole-array replacement:

• Omitting a field leaves it alone; null is how a caller unbinds — the shape set_text_range
  already uses for a run's bindings.
• Bindings apply after literals, so a bound field wins over a literal passed for the same field
  in one call.

Binding a font field has a prerequisite the numeric ones do not — measured:

bound field	needs a font loaded first?
fontSize, lineHeight, letterSpacing, paragraphSpacing, paragraphIndent	no
fontFamily	yes — unloaded font "<family> <style>" otherwise
fontStyle, fontWeight	not observed to, but same treatment

So the faces are walked as a chain in the order the bindings apply — family swap, then style swap —
and preloaded best effort. Failures there are swallowed deliberately: a guessed face may not exist
(a family from one mode paired with a style from another never co-occurs), and rejecting the write
for that would fail a binding Figma would have taken. Anything genuinely missing still surfaces,
from Figma, naming the exact face.

Ids resolve before createTextStyle(), so an id matching no variable creates nothing at all.
What that cannot cover — the target face is only computable once the style's base face is, i.e.
after creation when fontName was omitted — is undone with a remove() instead.

Measured limits

A multi-mode variable could not be tested: this account's plan caps a collection at one mode
(in addMode: Limited to 1 modes only). The preload takes every mode's value into account, so the
single-mode case is exact and the multi-mode one is covered by the best-effort pass rather than by
evidence.

Verification

• 10 mutations of the new logic, each confirmed to fail at least one test.
• Live, end to end: the fontSize + lineHeight + letterSpacing update that threw an hour
  earlier now succeeds; a fontFamily binding resolves the style's face to Roboto Regular with
  no unloaded font; a bound fontSize reads back as the variable's 42; null removes one
  binding and leaves its sibling; and an unresolvable id fails with
  create_text_style: variable VariableID:9999:9999 not found, creating nothing.
• Every commit builds and passes the full suite on its own.