But that is not really much different from instead dropping the baudrate with a factor 50 - your 100kbit/s loop will still melt down to <= 2kbit/s of actual data. But with lots of limitations and assumptions and lots of bruising involved.

And that burst mode still requires the master to hold the slave select enough time before the data transfer starts, that the slave have time to react.

If the slave does that burst inside a pin-change interrupt, then it will lock up all other interrupts (timers, UART, ...) for that millisecond. And the full ms may not transfer data - some of the time must be reserved for reacting to the slave select.

If that optimized assembler code is instead called from the main loop, then that main loop must guarantee better than 500us loop time always, to be able to get 500us of slave communication during that ms.

Whatever happens, the master must activate slave-select in good time. And the slave must do an best effort to detect it and then start the busy loop while waiting for real data. If the slave has a 100us jitter time to detect the slave select, then the master must be prepared to always hold slave select at least 150us before start of data. And the slave must busy-loop for 50..150us before any data will start to arrive.