nukt) is one of the very first tables called by the shaping engine. In most fonts, the nukta forms are created here. Glyphs however, breaks the glyphs with nuktas (like
nukta-deva) and lets GPOS position the nuktas later. So, when
cjct tables are called, the bases are still separated from their nuktas. A sequence like क ् ष ़ (u0915 u094d u0937 093C) ends up as
nukta-deva, which, I understand, is incorrect. The correct sequence should be
nukta-deva (since ssa with a nukta is not a ssa). Likewise for all other conjuncts.
This could have been prevented if a nukta-ed
ssa.nukta-deva) was formed in the
nukt table or the
cjct tables are made aware of the presence of a nukta.