One way to rotate a multi-term observable is to temporarily xor each of the terms onto an ancilla qubit, and phase the ancilla qubit, like this:
If that ZZ term is one of the observables of a logical qubit, this is applying a rotation to that logical qubit. (A more extreme version of rotating a logical qubit would be swap the logical qubit's value for a physical qubit's value, rotate the physical qubit, then swap back.)
If you look at the circuit for a distance 2 surface code, with an S gate added on the middle measurement qubit halfway through the cycle, you see it looks a whole lot like the start of the above ZZ rotation circuit:
Although the end of the circuit looks different, the S gate is in fact performing the same function as the phase gate in the ZZ rotation circuit.
By orienting the hook error the wrong way, the ZZ rotation that this S gate is performing lines up with the ZZ product that is the logical Z observable of the patch. And so it rotates the encoded surface code qubit.
Another way of seeing that it works is to realize that, by tracking a product of stabilizers prepared by reset gates through the circuit, you end up with a stabilizer equal to the logical Y of the surface code. All the data qubits were initialized in the X basis, so you would have expected to prepare logical X not logical Y. You'd need an S gate to rotate logical X into logical Y... which is exactly what the S gate in the middle of the circuit did to the logical qubit.
This screenshot from crumble is showing the Y logical observable forming from the product of resets:
Once you understand why it works at distance 2, and how to expand a patch from distance 2 to bigger distances, the remaining work is to realize that the distance 2 injection can largely be overlapped with the patch expansion such that everything finishes in 1 round instead of 2 rounds.
