One version of a "why" answer is to study the history of the system: I would recommend looking at this paper and references therein (Danielsson 1948; Dresher & Lahiri 2005; Fikkert, Dresher & Lahiri 2008, Lass 1992). How did the data change so that the noun is pérmit and the verb is permít – it's a puzzle.
Another kind of "why" answer is, what are the formal mechanisms that are used to generate the patterns? The first place to look is Chomsky & Halle (1968) The sound pattern of English. Essentially, they simply have different rules, ones sensitive to the difference between nouns and verbs.
Hayes adds to the mix a representational property of "extrametricality" which in the 1980 version makes little sense, it is just a diacritic property, kind of a "phantom zone" for phonological material. An improved theory of extrametricality was proffered, that "extrametrical" means "outside of the phonological word" (the P-word being a representational entity that mostly reconstructs the grammatical notion of "word" in the phonology).
It is correct that extrametricality (however conceived) is not necessary to account for the English facts, witness the SPE analysis which preceded the concept of "extrametricality" by a decade. Hayes puts the burden on representations because that was the trend in those days, alternatively it could be in the rule of quasi-rules – Optimality accounts don't rely on representations, instead they have a principle of "non-finality" which is, essentially, the same as Isdardi's "edge avoidance" parameter.
As far as I can tell, Idsardi's account does have a parameter to account for this, it's just not called an "extrametricality parameter". The general scheme of his system is that you place abstract marks on various levels of representation according to a formula "Place a ___ boundary to the ___ of the ___-most element" and you fill in right or left in each of the three slots – this LLR means "Place a left boundary to left of the right-most element – in /badigupofa/ that would give you /badigupo(fa/. Another parameter is "headedness" meaning that you project (only) the "L/R-most" element to the next level. The way he actually skips the final vowel in Macedonian where you have antepenultimate stress is to start with the parameter RLR (right bracked to the left of the rightmost element), hence vodeničar → vodeni)čar, eventually vo(deni)čar and make the leftmost element of the completely-parsed substring (deni) be the head, i.e. actually stressed. There are many ways to skin a cat in this system.
No analysis that I know of has completely avoided the need for a rule-based restriction on how you assign stress to nouns vs. verbs in English.
