The dot-product kerning trick is the bit I keep coming back to — encoding each glyph's edge profile as a bitvector and letting the overlap popcount decide the spacing feels like the sort of elegant hack that only falls out of genuinely understanding the problem. Also appreciated the honesty about the LLM-assisted parts; the cycle-count analysis is clearly hand-checked either way. I wonder how far the packed-font idea scales before the packing overlaps start costing more than they save.