I agree, assuming the recording is done well.
"Forget about off axis response. Fix the tonality or you will never be happy."
Personally I do place tonality ahead of spatiality on my list of priorities, but tonality was not the topic of this thread. In general I agree with the approach of fixing first that which matters most.
Imo your injunction to "forget about the off axis response" overlooks a vital aspect of tonality: Most of the sound you hear in most rooms started out as off-axis response.
You can EQ the response such that the sum of on-axis + off-axis = the tonality you desire, but if there was a significant spectral discrepancy between the two to begin with then it’s still there, and listening fatigue may arise over time. Let me explain:
The ear/brain system examines each incoming sound to see if it is a new sound or a reflection. It does so by comparing the spectral content to sounds recently stored in a short-term memory. If there is a match, then it’s a reflection and its directional cues are suppressed, but it still contributes to tonality and loudness. If there is no match then it’s a new sound, and a copy goes into short-term memory for comparison with subsequent incoming sounds. This suppression of directional cues from reflections is called the "Precedence effect", and it’s what allows us to reliably determine the direction of a sound source in a reverberant environment... useful for knowing where to look and/or where to run when a predator snaps a twig in the forest.
When there is a significant discrepancy between the spectral content of the initial sound and its reflection, the ear/brain system has to work correspondingly harder to make the correct match. Over time this can tire that portion of the brain and result in listening fatigue, sometimes literally manifesting as a head-ache.
One EQ-based way to minimize these spectral discrepancies might be to use a fairly directional (or possibly nearfield) EQ’d main array and a dedicated, separately-EQ’d reverberant-field-only array. Position and aim the second arrays (one for each channel) such that their outputs arrive after as much path-length-induced time delay as is reasonably feasible.
Duke