If the tristate buffer is so slow, I’m wondering what the result would have been of a AND at source and wide OR tree. The OR tree would act as buffers. The high latencies might make doing something like a SPI Target device hard (given read data has to be aligned against source clock, as opposed to xSPI with a source synchronous edge aligned read strobe generated by target device).
Because the spine is on `met4`, it's unlikely to cause an antenna violation, but to be extra sure, we have a protection diode on each input buffer wired to the spine.