odd nnrpd memory usage

[inn-workers at 2bb49e3f.mozone.net]

Maybe I'm missing something, but has anyone else noticed their
nnrpds growing substantially in size (FreeBSD 4.6-STABLE)?

On my box, configured with cnfs + buffindexed (yes, I switched 
back so we can try and nail that ovmap is null thingy), this is
what the typical nnrpds look like (tagged hash, yesterday morning's 
cvsup fyi):

%CPU %MEM    VSZ  RSS  TT STAT STARTED    TIME COMMAND
 6.2  4.6 190556 48564 ?? DNL   4:42PM 1:03.45 nnrpd: 10.0.0.3 XOVER
 4.7  8.0 190844 83360 ?? SN    4:43PM 0:59.11 nnrpd: 10.0.0.3  BODY
 3.6  2.9 190700 30792 ?? SN    4:43PM 0:51.16 nnrpd: 10.0.0.3  BODY
 3.6  2.5 190700 26424 ?? SN    4:48PM 0:43.09 nnrpd: 10.0.0.3  BODY
 2.4  4.7 190456 48740 ?? SN    4:48PM 0:37.54 nnrpd: 10.0.0.3  BODY
 3.2  0.7 190848  7824 ?? SN    4:59PM 0:30.12 nnrpd: 10.0.0.3  BODY
 2.8  0.5 190848  5004 ?? SN    5:08PM 0:11.91 nnrpd: 10.0.0.3  BODY
 3.4  3.5 190852 36104 ?? SN    4:02PM 2:26.79 nnrpd: 10.0.0.3  BODY
 3.8  1.8 190708 18940 ?? SN    4:37PM 1:07.80 nnrpd: 10.0.0.3  BODY
 0.0  0.1   2208   724 ?? INs   1:52AM 0:01.28 nnrpd: accepting connections

Most of these are in the *binar* groups, the pattern of the 
client appears to be XOVER->HEAD->BODY->BODY->BODY etc.  They
usually start off very small, similar to the 5004 RSS sized one
shown above, but then quickly grow up to 80-90megs in size. This
mem usage appears quite similar to the timecaf+tradindexed method
when I was using it a couple of weeks ago, however in that case,
both the VSZ and RSS's were close to 80-90megs each.

articlemmap is set to false, however changing it to true didnt
seem to make much of a difference at all.  Besides, technically
speaking, assuming worst case scenario that every object in a
given binaries group is maxarticlesize (~2megs in this case),
then the nnrpds ought not to exceed the base size + 2 megs for
each retrieved article, give or take.  To me, it almost apppears
like over time, the articles are just not getting freed.

Am I missing something obvious, or am I just not understanding
the way this is supposed to work?

nnrpds are spawned via: 'nnrpd -D -b 10.0.0.1 -n -s starting'

And, oh, the spool has about 8 1gig buffindices, 6 4gig cnfs buffs for
a class, and 6 ~110gig cnfs buffs for an alternate class.

-mohan