There seems to have been some discussion about this in the CR phase:
http://www.w3.org/2001/08/28-XSL-PR-DOC.html (see comment 20)
It would seem that in the case of reference-area generating FOs start-indent should simply be inherited (comment 20, item 3). In my example the start-indent of the block-container would be 10pt, which is then inherited by the orange block. This amounts to an effective 20pt indent for the orange block and oops, AltSoft is suddenly right. I'm confused.
I would say that in the absence of any clear right or wrong in the spec and different interpretations by different commercial renderers, I would suggest that implementing the easiest approach would be the best way to go.