There is a heck of a lot of complexity here, not the least of which is background theory.
First, you need to identify that a transform always happens between the scene referred domain and the display referred. Then you need to identify that the scene referred value set always needs to be aligned to your CGI attempt.
Given all of this, the proper solution is to not use a mirror ball, nor Blender. PFStools can generate linearized EXRs from cameras, but they will not be aligned for your scene, as you will have to scale the scene referred values properly, via strength for example, which is a simple uniform scale.
Also, more ideally is to capture your imagery via a lens, such as an 8mm fisheye. You can use longer focal lengths, but this will translate into more work stitching.
If you insist on trying to manually composite images in Blender:
- Import your footage linearized in 32 bit or 16 bit float ideally.
- Identify the near-linear response of your camera. It typically is a region below a nonlinear shoulder and above the nonlinear toe. Varies camera to camera.
- Use the "meatiest" runs of linear data per image, discard the rest. Fill in the missing portions with the meatiest portions from the other images. Assuming decent bracketing, you should have sufficient overlap that the transition between values is smooth.
This is effectively what PFS Tools does, so I would suggest to use it.
I would wager that the TIFF you are loading at 16 bpc is being loaded as nonlinear. Even if you flip it to linearized via the Properties in the UV, the result is still normalized to the display referred domain, so you would need to "decompress" it back to the scene referred values, in the context of your exposure for your scene.
Example: Exposure A at two stops down will have X number of pixels in the linear range. Exposure B will have Y. You will want to borrow the X pixels from A, and identify the same pixels in Y, which should have values very close to A after you multiply by 2^2, as the exposure is two stops hotter. Final image would be the portions of X composited with the portion from Y. Results will be scene referred, ranging from 0.0 to infinity.