DolphInNThe point of elo is to give an approximate skill level. The idea is that if you do better than expected, you gain elo and if you do worse than expected, you lose elo. There's reasons why we don't use simple W/L to adjust skill. Say that we have peak froyo vs a low invite team that play each other 1000 times, and froyo wins 5-4 for all of those matches. With the elo system, everyone's elo would eventually converge to reflect their skill levels, such as 2100 for the low invite team and 2200 for froyo. If we used W/L, the low invite team's lo would be deflated to below newcomer, and froyo's would be highly inflated. We used to have the winner's elo slightly decrease if they were expected to roll, but barely won, but enough players complained that we removed that, which is why we have the "elo out of thin air".
The problem with this argument is that you're assuming that the Elo number has literal meaning in itself, but really, Elo measures a difference in skill. The actual number isn't what matters; all that matters is the difference. If a team rated 1600 plays against a team rated 1400, the chance of winning for both teams as predicted by Elo rating is the same as if a team rated 2200 played against a team rated 2000. Elo is not a quantification of skill, but rather, a quantification of a measurable difference in skill.
In the situation you've described, you say that the low invite team that loses every time to Froyo would be placed in the "newcomer" Elo category. But the fact is, categories are simply labels you've placed on certain Elo ratings; if these two teams were the only ones playing, then the massive difference in Elo would reflect the difference in skill between the two teams, but it wouldn't say anything about the actual level of the two teams. Do you see what I'm trying to say? All that it would mean is that one team wins every time, and one team loses every time. The same Elo difference would happen if you had a Main team play against a Newcomer team over and over. Elo doesn't quantify skill; it quantifies a difference in skill.
What I'm trying to say is, there is no such thing as "invite" and "below newcomer" Elo in the example you give. All there would be is "top invite" Elo and "low invite" Elo. In your example, you're taking labels that you maintain out of one playerbase, that reflect that playerbase's relative skill levels, and applying it arbitrarily to a completely different playerbase. In a situation where Elo is based on rounds won, the same total divergence between the Elo of Froyo and the low invite team will happen, it will only be much slower (take many more matches played) than a situation where it is based on W/L only. This can be proven mathematically.
As long as Elo measures some zero-sum number, whether it be wins/losses or rounds won/rounds lost or something, then Elo will diverge to the same point, based on relative skill, no matter what. The only thing you're changing is the speed of the divergence. In a closed-box where Froyo plays a low invite team over and over, they will both end up in the same place whether or not you calculate Elo from round wins or pure wins/losses (assuming the score is the same every time, and assuming they play an infinite number of games).
The moment you start introducing non-zero sum elements into your "Elo" system, i.e., people gaining Elo without the other team losing the same amount, or thinking that Elo measures anything other than a difference in skill, you have committed the fallacy of objective value/a "quantification" fallacy. The only thing that would change in the example you give is the labels, since you're the one who creates the labels; the bottomed-out Elo would become the "low invite" rating, and the "inflated" Elo would be the "high invite" rating. Labelling different ranges of Elo is something you do once you see what Elo ranges various skill levels are in, and then name them accordingly. Thinking it's somehow bad/a failure of the Elo system that the team that gets beaten every time bottoms-out on Elo is like those people who used to say "I was Global before the rank reset!" in CS:GO. That is, you've made the mistake of thinking that a number can measure skill, rather than relative skill; you've assigned a meaning to the arbitrary name or "rank" (like those CS:GO players) that doesn't actually exist, rather than assessed the distribution of ranks in the playerbase.
Therefore, it doesn't matter whether Elo is calculated based on wins/losses or rounds won/rounds lost, so long as the amount of Elo gained by one team exactly matches the amount of Elo lost by the other team, so as to not violate the zero-sum nature of TF2/the nature of all Elo systems. Also, although it isn't technically necessary for a functioning Elo system (based on the above), it is super frustrating to lose Elo when you win. A much better solution is to simply award less Elo to the winning team for close wins and more Elo for stomps. This will achieve the same effect without the frustration.