Effects of restricting most significant bits

It should work for all channels, since they are all connected to the same ADC (it just switches between them)
You can confirm this by taking an offset reading on each channel to make sure they agree with each other.

LTC2485 is a "24-bit plus sign" converter. So, officially, there are 25 bits in 2's complement representation.

The 2's complement number is bits 30 to 0. The least significant 6 bits are not part of the 25 bits (sort of bonus garbage bits that probably have little meaning in most situations).

Bit 31 is more-or-less ** the complement of Bit 30 under normal conditions when the input is inside the normal operating range** (\$-0.5\cdot V_{REF} \lt V_{IN} \lt 0.5\cdot V_{REF}\$). If bit 30/31 are the same and the result bits are as shown in table 3 (all zeros with both high, or all 1s with both low), you know the input is larger or smaller than the reference voltage (overrange condition).

Oftentimes that would indicate a broken sensor or similar issue so you'd want to deal with it appropriately.

** except for a possible difference right at zero input.

So, your 25 bit signed result can be had by left-shifting the 32-bit number by one bit and masking out the least significant (now) 7 bits.

The least significant 6 bits are going to be pretty much all noise, but if you're doing calculations in double precision or 64 bit biased fixed point, it probably doesn't cost anything extra to use them (you can consider the result as a 31 bit signed number once you left-shift it). ENOB (Equivalent Number of Bits) of most ADCs is not better than about 20, bits and that's under ideal conditions.