I thought this was one of the better L&L articles to date. I do agree with the notion that the more detailed the rules are, the less wiggle room the DM has for making judgment calls. Its not so much that the designers are saying "Hey DM's, you can't change our rules!" Rather its the growing numbers of rules lawyers who are saying "Hey DM, according to page 137 of this book here, the DC is 10, not 15!"
While it is easy to blame it on rules lawyers as such, I actually think there is something much more fundamental going on. Rules are essentially defaults, and defaults are psychologically powerful. Even when they are acknowledged as suggestions, we often don't act like it.
Therefore I think it is good practice in presenting rules to state a global rule 0, but to also affirm it specifically in various parts of the rules even though this is, strictly speaking, superfluous. In such places this resets the default from something apparently rote to explicit permission to adjust as necessary, and can change our understanding of designer intent with respect to that rule. It helps disarm rules lawyers (yay!), but more importantly for many people it helps disarm our own reticence.
As for how does the barbarian become so knowledgeable in history, or how does the fighter learn so many details of the various religions? Well, the easiest explanation is that during their travels over the last several years, they actually picked up a little bit on what the cleric and wizard kept talking about around the camp fire each night. They still are not nearly as knowledgeable as their party mates, but some of it is going to stick over the years.
The high variance of the d20 really hurts this interpretation, in my opinion. I prefer training to be a strong predictor of success, but that means the change in typical results (usually from a bonus) have to be significant compared to the variance of the d20, and doing so quickly leads to inflation beyond the original range of the die. The lack of degrees of success (say +/- 5) as a standard principle also hurts. So a 30th level 4e fighter of average starting dexterity played as unsubtly as they come has better stealth than a trained first level rogue. To me that represents more than a little learning here and there. It's not just that the characters are better at some things than we would expect based on what they actually accomplish during the game, it's that they can't be bad at the things we've established them to be bad at. The player could ignore it and treat it like a 0, but then we're not really talking about the (4e) rules anymore.
Clearly this aspect of 3/3.5 was closer to what I like, but it failed to keep the huge bonuses reined in. If I were designing a 30 level 4e-style game from scratch I might make skill training be +3/tier, but you get fewer skills per tier. Perhaps 1 fewer skill at each tier, creating a sort of pyramid, but with class skills (e.g. Arcana for a wizard) automatically increasing. Fewer skills in the stratosphere makes them more special, and permits more typical DCs to be in a range where even marginal training is very helpful.
I'd probably also remove ability score increases from the game, so in general a +5 bonus is the best PCs can get there. As an ad hoc rule of thumb I'd maybe want a hyper-specialized epic character's worst result to be equal to the best possible result of an average untrained individual, which is a DC range of 1-40. In that case there is an additional +6 that could come from magic items, buffs, what have you. Enough, anyway, to feel significant if they showed up. Finally, I would concentrate on mechanics that don't increase the absolute potential of the character. Roll twice keep highest, for example, moves the average result closer to a character's potential. I like that for things like skill focus.