Unless there’s a bug, the behaviour you’ve described is not quite what happens.
By default the driver attempts to minimize the number of requests. The primary way this is accomplished is by spanning gaps between registers, up to the max register setting (default 125).
This means a request for {HR1, HR2, HR3} would request only HR1 through HR3 (3 registers), but a request for, say {HR1, HR2, HR10} would request HR1 though HR10 (10 registers) instead of using 2 requests to fetch HR1 + HR2 and HR10.
You can disable gap spanning entirely with the advanced “span gaps” property. You can also play around with lowering the amount of registers that can be requested at once, as you’ve found.
Either option is going to increase the number of requests required to get the data you’re asking for, sometimes dramatically.
If your data is distributed sparsely across the register range you’re going to have a bad time trying to reduce requests/traffic…