So I finally have time to test how a pair of relays behave when in series with one lamp.
There are many things to notice, most obviously that there is a strong hysteresis due to the contacts and the thermal capacity of the filament. Another thing I suspected is that all relays have their own threshold voltage and I'll have to build some "margin"...
I wired the relays in series so I can't actually tell which one turns on, but they need 7.5V to both turn on, and < 3.6V to turn off (that's a ratio of 2).
But these measurements were done at a very low speed, giving the filament time to heat or cool. The relays both turn on at 4.4V when their coils are switched instantaneously. However you need to wait 2 seconds for the filament to cool, before applying another pulse that works again.
A capacitor in parallel with the lamp can increase the frequency but I must see what I have in stock. I don't have enough electrolytics, I hope ceramics will do.
It looks like 5V might be a decent working voltage, a good compromise between speed and availability of the PSU.
I see already that the control pulses must have a low duty cycles. This should reduce the time to heat the filament, so the latch masters' relays should be "normally" connected to the input, and transfer the capacitor's charge to the output during the short pulse. This is because the capacitor will take some time to charge through the resistor. OTOH it will discharge instantly through the coil of the slave relays.
Now I have to measure the toggling speed I can reach at 5V...