Now being on my laptop rather than my phone, I can have a closer look and provide better observations, experiences and explanations.
But first,
Many thanks for the compliment. 
I try. FDM keeps me humble though as it has so many interconnected parameters and influences that, just when I think I understand it a little bit, something unexpected forces me to learn more.
AI seems to be a slow learner though. While I think this community finally managed to teach AI about the importance of drying filament, I am still waiting for AI to understand that fiddling with z-offset gcode to address a contaminated build plate is not such a good idea 
Especially when modern slicers allow easy checks using 1st layer line width 
Accepted and appreciated.
So looking at the pic’s on the big screen, my prior advise still holds. But the priorities change a little.
It was mainly the blue circled failures as well as my still vivid memory of my road to succesful dessiccant containers back when Bambu released the original PETG Basic (with, at the time, really bad default profiles) that led to me to point to curling first.
For a print like this, curling (a hot layer cools down rapidly on top of a cooler but still soft edge and overhang area) is a real threat as hot layer cool down can be very rapid while lower layers are not cool/strong enough yet.
Nevertheless, I now assume the first pic to really show a smoking gun (addressed by my last point above):
From the brim being attached, I assume that at least this print did have good bed adhesion (blue arrow).
Green arrows: Wide print, lots of pillars, overhangs significant but doable
Orange arrows: Narrow print, few pillars, overhangs significant but no worse than before
Besides curling, what I have experienced with these type of prints is that the “Slow down for better layer cooling” kicks in. If enabled, the print is slowed to achieve the layer time set in the “Max fan speed threshold”.
In cases such as this, it can however trigger the law of unintended consequences: By slowing down the print, it will not only give more time for the layer to cool.
It also keeps the nozzle in contact with the print for longer. This can inject much more heat into the tiny pillars. And with the part fan moving with the nozzle to the next pillar, the heat just stays and accumulates. So the effect of this parameter can actually be rather counterproductive. The Min print speed parameter is intended to counteract this somewhat, but that just forces the printer to limit its slow-down, bringing us back to the original problem “Slow down for…” was intended to solve…
==> You’ll probably want to increase layer time by printing two or more dessiccant containers at the same time.
The remainder of my advice above still holds. Although you may want to fiddle with “Wipe on retract” only if covering the pre-requisites, curling prevention and printing multiple containers at the same time does not give the desired result.
There is one more item of (minor) note. Although the consensus seems to be that filament color should not really have much of an influence, my past work involving nanocomposites have made me appreciate the massive effects that nanoparticles (like pigments) can indeed have on all kinds of properties. In particular, but not only, viscosity and hence rheology. I personally do have the impression that particular colors do print slightly differently: blue in particular but to a lesser extent black as well.
But I may be imagining this. Hallucination is not limited to AI .
& 
