PIE athematic neuters

Here's something that puzzles me: PIE athematic neuters seem to show an
regularly added extra mora in the vocalism of their nom./acc.sg., and it
isn't clear to me what kind of process may be responsible for this extra
length. I know that Jens explains some of the forms in question as old
collectives with final *h2, but I have several reservations concerning
this explanation.

First, the structure of the neuter stems where the lengthening occurs is
usually CVCR-/G (G = glide) and I find it difficult to believe that the
lengthening caused by *h2 would have operated across this type of
cluster, e.g. in the acrostatic type of ***de:rw-h2 > **de::rw > *dóru
(weak stem *déru- < *de:rw-'). My impression is that the actual regular
collectives are those like *wédo:r (beside *wódr., gen. *wédns).

Secondly, the wholesale survival of collectives and the equally
wholesale loss of non-collectives at the same time seems an odd thing to
propose. What happened to the "plain" words for 'water', 'tree', etc.?

Thirdly, we see the same thing e.g. in acrostatic adjectives like
*h1we:s-u-s (genus communis) vs. *h1wos-u (neuter) (judging from the
forms that have survived analogical levelling). Note that here the
nom.sg. *-s of the common gender fails to change the quality of the root
vowel!

The extra length occurs in the following types:

*k^e:r(d), *je:kWr. (preserving their long root vowels), and possibly
all proterokinetic and amphikinetic neuters. The fact that neuters are
never hysterokinetic is striking. The root vowel of athematic neuter
stems _systematically_ attracts accent and resists reduction in the
nom./acc.sg., which may point out to original length.

The acrostatic type represented by *doru, *h2oju, *g^onu, *h1wosu (weak
stems *deru-, *h2aju-, *g^enu-, *h1wesu-), where the *o supposedly
reflects underlying _overlength_ (whatever its phonetic realisation),
always shortened in this type (presumably at a time, in pre-PIE, when
the final sonorant was not yet vocalised and so the long vowel was
followed by a consonant cluster). Here we also have such *-r/n-
heteroclita as *wodr. and *potr., which may be ancient *-nt-
participles, i.e. *we:d-nt- etc., with a long vowel in the verb root
plus another mora yielding the attested *o.

Any ideas?

Piotr