Good questions, and I like the money-burning idea: it's honest about what it does.
Feature vector escaping: nodes never store the raw feature vector. The phone sends it over an encrypted link, and the node turns it into a protected (cancelable) template and keeps only that. A stolen template on its own is scrambled numbers. As detailed in the whitepaper, today the key that scrambles them is held by the full nodes (currently the project's own servers), so a compromised node could unscramble its templates back into feature vectors. Never a photo, but a feature vector could in principle be compared with the same model run on a photo of the same person. That is why rotating the key and splitting it across independent operators are next on the roadmap.
Surveillance-video spoofing: enrolment needs a flat palm close to the camera plus a live face check where the app asks for a random movement, which ordinary surveillance footage does not give you. Currently, that liveness check runs in the app, meaning a modified app could theoretically bypass it. Server-verifiable liveness is the next item on the roadmap. What the nodes do enforce already: a palm or face that matches someone enrolled is refused, so a person cannot open a second account with their own body.
Your scheme versus ours: yours makes abuse expensive, which works for spam. We went for one account per person regardless of money, because the same accounts vote on the network's rules, and with a price per account whoever can pay most gets the most votes.
(I'm one of the developers.)