There is no longer a requirement that all the masters have exactly the same base color glyphs as the default masters. Similarly, it's no longer required that all masters' LayerLists have the same total count of layers. It is sufficient that, for a base color glyph in the default master, a non-default master may (or may not) contain one with the same name and same effective number of layers (which in turn can be laid out differently in the respective LayerLists).
This provides greater flexibility when working with variable font project with sparse glyph sets.
previously we only reused the VarIndexBase of a previously seen variable table when the current's varIdxes were _fully_ equal to one of the previous; now we also try to find a match anywhere in the accummulated list of self.varIdxes, including a partial match at the tail of the list.
When multiple variable tables refer to the same delta-sets they can now share the same VarIndexBase so the resulting DeltaSetIndexMap is a bit smaller.
For simplicity, we only reuse VarIndexBase when variable tables fully share (ie. same, and same number of) varIdxes; potentially we could reuse subsets of varIdxes (e.g. a VarColoStop.Alpha has a +0.5 delta, and later on elsewhere a PaintVarSolid.Alpha has a similar +0.5 delta; the latter could have a VarIndexBase that reuses an existing DeltaSetIndexMap entry for the former), but for now this I think is good enough.
Before we were too greedy in the way we converted subtables to VarType. E.g. If a PaintTransform wrapping a PaintRadialGradient contained variations in the Affine2x3, we would incorrectly convert also the gradient's ColorLine (and ColorStops) to VarColorLine, VarColorStop, etc. (even if the gradient was not variable!). Instead we want skip traversing a given subtable including its children when the predicate doesn't match.
Introduce three compilation modes:
1. Pure python: only uses the existing fonttools packing and overflow resolution.
2. Harfbuzz+python: uses harfbuzz packing and python overflow resolution. Extensions are allowed to be shared.
3. Python fallback: if harfbuzz+python runs out of resolution options, this disables extension sharing and only uses python packing. Once it succeeds control is passed back to the harfbuzz packer to produce the final packing with extension sharing enabled.
this will be useful when computing deltas for COLR table fields that use Fixed or F2Dot14 (ItemVariationStore internally treats fixed-value floats as 16 or 32-bit integers)
though we still emit the empty <Map index=.../> elements; they help identify chunks of delta-set indices, and define the length of the array upon loading from XML
When building COLR masters, layer reuse may lead to different number of layers between masters, becasue some master may reuse specific layers while others may not. Add a flag to disable this optimization; will need to be run again after the VF COLR font has been merged
This does two things:
Fixes forced-set computation, which was wrong in multiple ways.
Debugged it. Is solid now... Famous last words.
Speeds up DP time by limiting DP lookback length. For Noto Sans,
IUP time drops from 23s down to 9s, with only a slight size increase
in the final font. This basically turns the algorithm from O(n^3) into
O(n).
The exception UnsupportedFormat was defined and then redefined with the same name in varLib.errors, and imported twice from varLib.merger, probably as result of a sweeping find/replace.
Rename it 'InconsistentFormats' as originally intended.
