Fixed for Deva cases with two full-form consonants.  Failures **way** down.
Not much left to go :-).

We now check that the 'rphp' table exists before forming Reph, so
we don't need to comment out Ra for those scripts.

Hindi Wikipedia failures down to 6639 (0.938381%)!

The spec says: "[{M}+[N]+[H]]", and that's what Uniscribe implements.
We instead do: "{M+[N]+[H]}", which means we allow Nukta and Halant
after all Matras, not just the last one.  It makes more sense.

Uniscribe doesn't do it, but we want to do as it gives the Reph the
opportunity to interact with the Matras.  Test with mangal for example.
Sequence: <0930,094d,0915,094b,094d>
In test suite already.

This was removing the mandatory global 1 bit in the mask and hence
disabling GPOS for Reph!

Move its dependents too.

In the case of Consonant,LeftMatra,Halant, Uniscribe leaves the Halant
where it is, but we want to move it with the Matra as that makes more
logical sense.

This introduced a failure, which we tracked down to a test case like this:

  U+092E,U+094B,U+094D,U+0930

The final character is a Ra that should be put in a syllable of it's
own.  And we do.  But it will interact with the Halant before it.  So
now we finally are convinced that we have to limit features to syllable
boundaries.  That's coming after lunch!

Reusing the consonant logic!

20 was way too low, one could hit a syllable with 7ish consonants with it.

Another 1.5% down.

Another 12 thousand failures gone! (78 to go)

Like Uniscribe does.

Shaping failures on Hindi Wikipedia go down from 25% to 14%!

Fixes 10% of the failures against all of Hindi Wikipedia!