One point which might help to understand soft-forks a bit better:

We have two class of full nodes active in Bitcoin mainnet, mining and non-mining nodes, the latter group, nodes which have no part in setting up a block candidate for mining devices to try their luck finding the magical nonce, could go back in the version number almost down to the original Satoshi client, while the other group, nodes that actively generate raw blocks for mining purposes has to upgrade almost to the latest version (forget about the reason behind using the word almost in this context, it doesn't matter).

This way, a soft-fork like segwit is not to be even sensed by non-mining nodes as long as they are not interested in segwit news like who has how many coins in her bech32 addresses, is this segwit utxo spent or untouched, etc. The trick is transferring coins from legacy addresses to segwit ones is seen from the perspective of legacy nodes as a weird ANYONE_CANSPEND utxo which is never seized by anyone, as miners have migrated to segwit, and they do not put such seizing transactions in the blocks they generate.
 
Interestingly, miners were not the main supporter of segwit, it was somehow enforced by a majority of non-mining nodes aka users, they would fork out if a block comes with a stupid seize of segwit locked utxo. It was a bit more complicated situation compared to my simplified explanation, twist that is called User Activated Soft Fork, UASF.