Also my scheme is nothing like Firstbits, because it does not depends on centralized trust.
FirstBits didn't require centralized trust. Anyone could using the algorithm and a copy of the blockchain determine the firstbits of an address or the address from a firstbits. It is no more or less centralized then the scheme you proposed. It is more efficient but no more secure. If you compute it yourself it is secure and if you don't you are trusting the person who computes it.
Telling them "don't do that because you put others at risk" is not an option. If this is the only downside, I'll implement my scheme with a fixed Bitcoin address and just wait better support for stealth address + a lift of OP_RETURN.
Well that isn't the only downside. What happens when someone who uses a wallet that doesn't support this scheme sees the mnemonic and wants to send funds to it. How are they going to convert "yellow happy dog runner" into the proper address? Well they could download some additional open source software, make sure it is legit, and use that along with the blockchain to compute the proper address. They "could" do that but lets be honest they won't. We already they won't because they could have decoded firstbit addresses on their own but they didn't. What they will probably do is google "name-of-your-scheme" and very likely find a dishonest site which will decode "yellow happy dog runner" to an incorrect address owned by the site operator. Maybe they even show the wrong transaction in the blockchain with a link to blockchain.info to make it look more legit. They send funds to the address they think is right, the money is gone and your user and the sender are both confused as to what happened.
So yes if 100% of wallets natively supported your scheme it wouldn't require any third party trust (beyond trusting the wallet developer which is implicit to some degree). Of course if 100% of wallets supported firstbits it also wouldn't require third party trust but we know how that turned out.