Thanks for the photos, context is helpful.
Before I did anything, I’d do range checks at several different bearings. See discussion on antennas and multipath towards the end.
Without being able to poke at it , all I can offer is a shot in the dark based on no particular insight: I don’t know if you have a sacrificial unit, but you might try cutting the PCB out from under the module antenna section.
Were both boards physically mounted the same way when doing the A/B comparison? Mounting can make a big difference in stray radiation and this can help or hurt. If the whole board is flapping at 2.45 GHz…
Maybe not a factor in this case, but with RF, ground never is. If you have two planes, the only way to get them close to each other is with generous via stitching. I’d try it in the next rev and make it a habit with RF. The higher the frequency, the closer the vias. As a thumbrule you want stitching WAY under a tenth of a wavelength. At 2.5 GHz this is1.2 CM, BUT the dielectric constant of the board makes a wavelength in board about 60% of this or 0.72 CM, but call it 5 mm. And this is just a thumbrule. For stuff like coplanar waveguide or low impedance RF stuff like PA output circuitry, lots more is better.
Another issue is that we tend to think of these radios as having a spherical wave front - equally good (or poor) in all directions. Reality usually departs from this in substantial ways. Antenna patterns are generally messy, with peaks and deep (20+ dB) nulls. A lot of engineering effort is applied to controlling patterns. I would not be at all surprised to see dramatically different ranges on your range measurements on slightly different bearings. At these frequencies reflections are a major concern. The environment is a lot like looking at a flashlight in a hall of mirrors. This is the big reason the stuff works reasonably well in spite of poor antennas. And that reflection environment (RF Systems types call this multipath) changes in real time. Does your unit have a cooling fan? It’s a moving reflector. Plastic blades - moving dielectric (think lens) still causes changes. Some one walking by - looks like a sequined Vegas showgirl to BT.
What you describe is common. The only reason this works at all is the multiple paths caused by reflections. It’s likely the metal box is pretty leaky at 2.45 GHz unless a lot of effort was taken to make it RF tight. I’m betting no effort was taken. Still, it’s going to have some impact, and it’s mostly going to be negative.