1. I have solved my practical problem. I still don't know
what exactly happened before -- I don't care. As a hint:
It has worked since I have run \outputpenalty=-10004
with \holdinginserts>0 (I reported before that otherwise
the \footins box isn't empty with the ensuing
\outputenalty=-10002, in spite of \holdinginserts>0).

2. With simply switching \holdinginserts as sketched
before, you create another bug (well, wait:). \@specialoutput
measures the \footins box ..., so in adding floats it doesn't
take account of any footnotes after \holdinginserts>0.
I have made an example giving an awkward outcome
-- guess: a float that better should have waited for footnotes,
but one may wonder how real-life this is, I am preparing
something perhaps more digestible, may be for CTAN or
what someone suggests ... however, this seems to favour
what Frank Mittelbach (at least some frank from latex-bugs...)
expressed in latex/3370: the rewriting will be done for LaTeX3,
not 2e.

-- Uwe Lueck.