I thin you're right to bring up this point. As far as I know, there are alternative ways of looking at the rules for English stress and vowel quality/reduction, and I don't think the matter has been definitively settled yet.
One thing that hasn't been mentioned in the data that you give (where did you get it from, by the way?) is the apparent relevance of morphological structure: astonish ends in -ish, which can be identified as a verb suffix or at least as an ending that shows up on a number of verbs (consider pub-lish, pub-lic; pun-ish, pun-itive and fin-ish, fin-al for some possible evidence for the existence of -ish as an identifiable suffix, although none of these are productive patterns), whereas improve, combine, believe, collapse, exist, resent, revolt, insist all start with prefixes (from an etymological perspective, at least).
In some prefixed disyllabic words, the surface form ends in CVC but the second syllable is still stressed: emit, omit, reset, refit, repel, dispell, compress, impress, redress. These can be analyzed as having an "underlying" form ending in a "virtual" geminate (see the previous question English stress, abstract analysis), but this makes the situation less simple than the originally stated rule implies.
Unless we adopt a "virtual geminate" hypothesis (or something similar), this class of words would also pose a problem for your alternative hypothesis that "The vowels may as well become long because they're stressed".
One paper you might be interested in reading is "Rethinking English vowel alternation", by Jonathan B. Alcantara (1994).
