You're disregarding the second half of the sentence after your red box:
"and we assume the existence of a source truth that is prohibitively expensive for typical usage but is available in case of disagreement"
If a node flags collusion in Tier 1, and then the Tier 2 Oracles deem the report to be a false claim (when in fact it is a correct check against collusion), enter source truth to clarity dispute at the expense of the incorrect party. It's intentionally expensive to discourage disputes. So on top of paying for incorrect reporting, a collusion effort would also effectively ban your participation from future reporting.
Theoretically, your hair dresser could take off with advanced payment if they felt so compelled to do so, but that decision will impact their prospect of getting another hairdressing job in the future. The company will undoubtedly have a claim filed in the employees name, employee will be fired and won't have a reference needed at the next interview.
You dismiss this concept as unrelated to cryptonomics, and in the technical definition, you might be right. Practically, however, this is related to the incentive to act honestly. The cost of reputation damage that effectively bars you from participation far outweighs the chance of pulling off an elaborate scam unnoticed, with the best case scenario being hefty penalties for false reporting, and the worst case being banishment from participation altogether.
I think this argument is a bit reductionist. You're simplifying a complex system of checks and balances that has heavy incentives, both positive and negative, that become more magnified with economies of scale. More participants, larger contracts, larger payments = even higher incentives for both correctly reporting and monitoring.
Is it perfect? No. Is it pretty damn good? Yes. Will it get better? Definitely.