Cross-activation, proposed by Ramachandran and Hubbard in 2001, starts from an anatomical coincidence: the cortical region that processes written characters sits next to the color region V4. Extra wiring between neighbors, perhaps from reduced pruning in development, would make letters light colors automatically.
The theory predicts that common forms should link neighboring regions, and grapheme-color synesthesia’s dominance fits. Diffusion imaging has found increased structural connectivity in synesthetes, which is consistent without being decisive.
Its main rival, disinhibited feedback, needs no extra wiring at all. Current evidence suggests the boring, likely answer: different forms may use different mechanisms.