The short answer is that you should not use an explicitly modeled Coss in conjunction with a PLECS thermal model as it invalidates the assumptions of the thermal model - namely an instantaneous transition from the off state to the on state.
You end up chasing various artifacts 1) the instantaneous current discharge of the Coss through the switch on-state resistance and 2) the parallel capacitance means the voltage across the switch voltage cannot change instantaneously at the switching moment. These distort the inputs into the look-up tables. You can use the workarounds you mention in your post like using the tank current, perhaps eliminating the artificial power losses from discharging Coss, but this isn’t a consistent engineering workflow one can make broad recommendations on.
The typical recommendations are twofold:
- Clearly separate the thermal model used for loss analysis (requires some assupmtions, for example ZVS is always achieved or ZVS losses) from the model used to analyze ZVS/dead-time commutation. This is explained further in this post.
- Use PLECS Spice which, with the appropriate model of the semiconductor device, can explicitly model the Coss, ZVS behavior, and power loss.