variationModel.interpolateFromMasters doesn't extrapolate as expected

[LettError]

With extrapolate=True there is an issue withvariationModel.interpolateFromMasters as it clips results in 2 value models. Results are as expected in 3 value models.

from fontTools.varLib.models import VariationModel

# extrapolate under and over a 3 value system
locations = [
    dict(wght=-1),
    dict(wght=0),
    dict(wght=1),    
]

values = [30, 40, 50]

m = VariationModel(locations, extrapolate=True)
assert m.interpolateFromMasters(dict(wght=-1.5), values) == 25
assert m.interpolateFromMasters(dict(wght=-1), values) == 30
assert m.interpolateFromMasters(dict(wght=0), values) == 40
assert m.interpolateFromMasters(dict(wght=0.5), values) == 45
assert m.interpolateFromMasters(dict(wght=1), values) == 50
assert m.interpolateFromMasters(dict(wght=1.5), values) == 55


# extrapolate under and over a 2 value system
locations = [
    dict(wght=0),
    dict(wght=1),    
]

values = [40, 50]

m = VariationModel(locations, extrapolate=True)

# values under 0 are clipped.
print(m.interpolateFromMasters(dict(wght=-1), values))
print(m.interpolateFromMasters(dict(wght=-2), values))
print(m.interpolateFromMasters(dict(wght=-3), values))

#> 40.0
#> 40.0
#> 40.0

# this one fails
#assert m.interpolateFromMasters(dict(wght=-1), values) == 30

# the rest works as expected
assert m.interpolateFromMasters(dict(wght=0), values) == 40
assert m.interpolateFromMasters(dict(wght=0.5), values) == 45
assert m.interpolateFromMasters(dict(wght=1), values) == 50
assert m.interpolateFromMasters(dict(wght=1.5), values) == 55

[anthrotype]

at least it appears to be intentional e.g. see the docstring of VariationModel (emphasis mine)

5284ac3

If the extrapolate argument is set to True, then location values are
interpretted [sic] in the normalized space, ie. in the [-1,+1] range, and
values are extrapolated outside this range.

I'm sure @behdad can clarify that.

[behdad]

What @anthrotype said. Current extrapolation implementation extrapolations the values at -1 and +1.

[LettError]

Would it be possible to express the hope it would eventually also work for values between [0,1]?

[behdad]

I have to check the code. I think it's possible, yes. In that case it wouldn't need any specific range and would adapt to the min/max specified.

[behdad]

This is trivial for one axis, less so for more axes.

[LettError]

True. In MutatorMath that was a bit of work (lines starting here). This is too specific and can't be re-used for varlib. But I imagine it will be similar enough: a couple of different cases that need to be calculated separately.