I take it you are converting from EPSG:4283 (GDA94) to EPSG:7844 (GDA2020), or between their sibling codes (4989 and 4938 for the former, 7843 and 7842 for the latter), correct?
By default, datum conversions go through the common base datum - WGS84, the realization used in EPSG:4326.
EPSG has a single applicable transform from 4283 to 4326, that's transform 1150, 'GDA94 to WGS 84 (1)'. The parameters are all trivial, but the accuracy is 3 meters, so the transform says that 4283 is equivalent to 4326 with the error of 3 meters. There is a note for the transform: 'Approximation at the +\/- 3m level using inappropriate assumption that GDA94 is equivalent to WGS 84. Accuracy changed from 1m to 3m due to tectonic plate motion over more than 15 years.' The assumption used in the transform is called to be inappropriate, yes, but the transform is in the system, it is not marked as deprecated, and there are no other transforms to WGS84 with better accuracy.
Same for 7844, EPSG has a single applicable transform from 7844 to 4326, that's transform 8450, 'GDA2020 to WGS 84 (2)'. The parameters are also trivial, the accuracy is also 3 meters, there's also a note: 'Approximation at the 3m level assuming WGS 84 is equivalent to ITRF2014 within the accuracy of the transformation. See GDA2020 to WGS 84 (G1762) (1) (code 8448) for a better approximation and ITRF2014 to GDA2020 (1) (code 8049) for actual relationship.' Again, the transform is said to be not terribly good, but it is in the system, it is not marked as deprecated, and there are no other transforms to WGS84 with better accuracy.
That's why the default conversion between EPSG:4283 (GDA94) and EPSG:7844 (GDA2020) does nothing. Maybe we should reword the synopsis from 'coordinate systems are equal' to something else, although I couldn't come up with good choices quickly ('no conversion' would be bad because it might suggest that there will be an error).
There are, however, direct transforms between the two systems which we find and allow selecting from, as you said. These are way more accurate and more appropriate.
We don't use direct transforms by default because (a) there are sometimes multiple direct transforms between a pair of systems with no automatic choice possible, one transform is better for one geographic area and a different transform for the other, and (b) there are frequently dependencies on grid files which might not be there, they are an optional install, bulky, are updated on a different schedule than the builds, etc. If we start doing the 'if the grid file is there, let's use it, otherwise let's use the default transform' we are back to having the default transform sometimes not being very accurate except it got harder for both the user and us to understand which exact transform is getting applied to the data. So, the default is always the best transform to WGS84 (best accuracy among those not deprecated) which does not use any optional data so that it is the same for all installs of the same build. In this case this ends up having the accuracy of 3 meters, unfortunately.
Sum total, maybe the language should be clearer but the behavior is expected.