I have created elliptic curve VRF for TON: https://github.com/ProgramCrafter/ecvrf-coordinator-contract.
Its coordinator contract in testnet is kQAkRQPZL7siQjoc26FFXRGgb/r/ZzmdP4P+anwZfcwvkWkh, accepting messages from users according to the following TL-B:
op::subscribe_random#AB4C4859 consumer:MsgAddressInt = InMsgBody;
op::provide_random#069CECA8 rnd:u256 = OutMsgBody;
Script calculating the random values and sending them to blockchain is private but it's not ...
You can't return these jettons unless @wallet team manually does that.
Actually, that's a problem of custodial services and reason why you can't call yourself "the owner" of the wallet there.
In testnet environment, there is opcode GASCONSUMED, doing exactly that. The update isn't shipped in mainnet yet.
It should be possible to cut off the last half of key (64 HEX characters), which is actually public key, and use the rest in other wallets.
Strictly speaking, finding the most efficient way to structure storage is NP-class task meaning that it will require too much resources given the amount of bits.
Dictionaries are binary trees (up to 2 branches in cell), though reading and updating them is actually cheaper than quadro-trees. So I'd recommend using dicts.
Yes, the owner will eventually lose all the money.
However, data stored by wallet (let's assume v3r2) is less than 1 KiB, so to lose even 1 TON you'll have to wait thousands of years, over which time the network will clearly have some other breaking changes.