Hi @Krzysztof.N ,
Thank you very much for a so detailed reply. I don’t know how to quote you partially:
I have tested multibounce, working great now, thanks! 
It’s looking quite near to old RT reflections, so I’m quite happy. As weak points, as you said, it’s super noisy for mid-rough surfaces, as a polished concrete. It’s even better to not use temporal accumulation for reflections, as it creates very ‘fat’ noise. (I’m referring to the second bounce reflections, the ones in the mirror).
Also, I have noticed increasing the max roughness is a performance killer, while in RT reflections doesn’t have that big impact.
Temporal (and it’s worse in realtime):
Without temporal:
About the emissives, thank you, I already tested those workarounds and published one of them here: Possible to turn off Emissive Material for Lumen GI? - #22 by Miguel1900
Also tried reducing the cvar intensity to 10, but it’s still there. I would need to reduce to 1 but it’s a huge tradeoff.
Based in your explanation, I think it may be a bug, as it has a similar behaviour when “normal” or when ray traced:
(Much more noisy in realtime)
This persistence of light in the last one (the light propagation, so the noise too, dissapear when out of screen space) seems related to r.Lumen.ScreenProbeGather.ScreenTraces.HZBTraversal.HistoryDepthTestRelativeThickness and/or r.Lumen.ScreenProbeGather.ScreenTraces.HZBTraversal.RelativeDepthThickness, which I had with a value of 0.1 instead of default 0.01. The higher it’s, the most splotchy. But as it’s not 0 by default, it will be a little noisy with the default value too, without the possibility to remove it at 100%.
And thank you for the rest of the reply! I also understand the commercial point of view, but I was affraid of having thought about “bread for today, hungry for tomorrow” (The possibility of Lumen scalabiliy being limited due to a nowadays-technology thinking, instead of a future projection).