Incorrect bone placement on some FBX animation files

Here is animation which displays an incorrect placement on some bones (specifically the Clavicle ones) when the animation is exported with the model itself (Anim_walk_loop). When the geometry is not included in the animation file (Anim_walk2_loop.FBX), the problem disappears. We identified that this is caused by Position keys missing in the generated HMD file. When this optimization is disabled, the problem is resolved:

https://github.com/HeapsIO/heaps/blob/dd0f74e7c00997c70c2a234285027c55b04c75a7/hxd/fmt/fbx/BaseLibrary.hx#L989

            if( data.z != null && roundValues(data.z, def.z, M) )
                hasValue = true;
            // no meaningful value found
            if( !hasValue )
                continue;      // <-----------
            var keyCount = 0;
            if( data.x != null ) keyCount = data.x.length;

Sample file:

NG_anim_bug.zip

The problem is correctly understood.

We're trying to know if the position in the animation is different than the position in the default skin (in that sample there's a difference for L/R Clavicle)

if the position is animated (non constant) then we will force the position into animation data
if the position is not animated (constant), we will compare it to the skin and force if different.
if there is no skin (no model, like the good animation in the sample), we are forcing the position as well (if( !hasTrans && def.transPos == null ) in BaseLibrary.loadAnimation)

The bug occurs then when:

we have a model in the Animation FBX
the position is not animated
the (fixed) position is different than the Model one

The correct fix would be to either:

correctly deduce the original skin position (defMat) : I tried to do that (commented code in BaseLibrary.updateDefaultMatrix) but it didn't work very well. Also would require to have skin (not work if model not exported)
always force position for all animations, whereas skin is present or not. This would increase animation size. It would be better if we can separate T/R/S data so we can have "PerFrame" on a transformation basis and not for the actual object

HeapsIO / heaps

Incorrect bone placement on some FBX animation files #1115