Perhaps you could even stipulate that the new seed phrases must be either 15, 21, or 27 words long, and so any seed phrase which is 12, 18, or 24 words is immediately known to be a legacy seed phrase.
Regarding false positive rates, an 8-bit version field would result in an
nVersion/256 false positive rate.
Assuming support for three versions, this would yield a false positive rate of less than 1.2%.
If this isn't enough, one option is to use the 24-bit general purpose field to mitigate this.
This has already been addressed by Lukechilds here :
https://github.com/lukechilds/bip39-versioned?tab=readme-ov-file#false-positives