Some of the highlights included non-maskable interrupts ignoring the state of the IFF1 flag (this flag is automatically cleared when an interrupt is serviced, and is used to prevent the interrupt handler from being called again before it has finished) and the RETN instruction not copying the state of the IFF2 flag back to IFF1. When non-maskable interrupts are serviced, the state of IFF1 is copied to IFF2 before it gets cleared, the idea being that if you use RETN interrupts are automatically re-enabled on exit of the NMI ISR. (Contrast this with maskable interrupts, where both flags are cleared, and you need an explicit EI to re-enable them).
The HALT instruction (executes NOPs until an interrupt is requested or the CPU is reset) was also completely incorrectly (and bizarrely) implemented. The rewrite just sets a Halted property, which prevents the CPU from fetching or executing any instructions. The interrupt-triggering code simply resets this property.
This has fixed numerous bugs (I'm not sure when they were introduced, as it was all working a while back). It's gone from "not working at all" to "just about working", but some games or demos that rely on precise interrupt timing don't work properly.
Game Gear Hicolor Demo and Desert Speedtrap
Both problems in the above screenshots relate to line-based interrupts from the VDP (Video Display Processor). Some other games simply hang at startup. [sad]