This is a not-so-well-documented behaviour
of Table Calculations when At the Level is in play.
In your case there is another layer of complexity
because of Data Densification (or rather the absence of it) --
please check the # of Marks on the heatmap2 view (784)
as opposed to the other ones (740 -- no densification).
Luckily, there is the man who wrote the Bible on the very topic --
calling Jonathan Drummey and referring to his 2-Part article:
To add my 2c, please find the attached with Sheet 7
(using the recipe from the Part 2 of the above).
trellis_9_3_YF.twbx 1.4 MB
Thanks a lot for the help. I had read through Jonathan's in depth discussion before on level of detail, but missed the part at the end of the first article where he explains how domain padding acts differently when dimensions are on the detail shelf instead of rows, columns or pages. That turned out to be the critical distinction here.
Here is the relevant quote from Jonathan's article:
"if the dimension used for At the Level (or any addressing dimensions above it) are on the Level of Detail Shelf while other addressing dimensions are on Rows, Columns, and/or Pages, Tableau triggers domain completion for those addressing dimensions."
"Another [way] to prevent unwanted domain completion is to put all your addressing dimensions on the Level of Detail Shelf, and use discrete measures such as ATTR(myAddressingDimension) to generate the proper headers."
A well respected zen master responded that he now tries to pad out his data in advance of using Tableau to avoid these kind of domain padding issues.
This is the diagram that made the behavior (almost) clear to me at the end.