Looking at the unmodified schematic (pages 5 and 6):
https://github.com/myriadrf/LimeSDR-USB/blob/master/hardware/socket/1v4/LimeSDR-USB_1v4s.pdf
The matching networks, which are what I would call front end filters, are optimal (by default, no modification) for:
LMS7002M RF1
TX1_1: 30MHz to 1.9GHz
TX1_2: 2GHz to 2.6GHz
RX1_L: 700MHz to 900 MHz
RX1_H: 2 GHz to 2.6 GHz
RX1_W: 700MHz to 2.6 GHz
LMS7002M RF2
TX2_1: 30MHz to 1.9GHz
TX2_2: 2GHz to 2.6GHz
RX2_L: 700MHz to 900 MHz
RX2_H: 2 GHz to 2.6 GHz
RX2_W: 700MHz to 2.6 GHz
The below links show a possible hardware modification to RX1_L or RX1_W/RX_2_W
LimeSDR - Optimising Performance on HF and Optimising LimeSDR matching for HF – MyriadRF
The tuner can technically go from 100 kHz to 3.8 GHz, but the amount of attenuation on each input will vary as a function of frequency. It would be really good if Lime, or anyone else with the required hardware, could provide frequency response graphs from 100kHz to 3.8GHz on each of the 6 RX inputs (RX1_L,RX1_H,RX1_W,RX2_L,RX2_H,RX2_W) as the input frequency is swept from 100kHz to 3.8GHz with a constant input level of say -45dBm.
Looking at the above information from the schematic I would assume that on an unmodified board RX1_W and RX2_W or RX1_H and RX2_H should provide the least attenuation for 2.6 GHz to 3.8GHz The filters look like first order (or possibly second order), so the roll-off should be gentle. If it is first order then it has ~6dB per octave roll off. If I assume that it is a third order filter (~18dB per octave roll off) and that the attenuation was 10 dB outside the pass band (@3.8GHz), I’ve no idea what it actually is, but a theoretical 12-bit ADC has a dynamic range of 74 dB, so even if the signal of interest was being attenuated by 10 dB you should still be able to see it, if you are not then you need to look at adding external pass/block filters and external gain. While making sure that all signals entering the LimeSDR are less than +12dBm, I would be aiming for 0dBm, or even -10dBm, to give plenty of headroom. But I do not think that is something that you need to worry about if you are just looking at 2.6 GHz to 3.8 GHz, because the roll off is so gentle.
Not an easy question for anyone else to answer, only you know how you plan on using your LimeSDR. For me leaving one input unmodified, which will filter out more noise outside the pass band makes sense at least for me.