There are a few motivations here:
1) Less boilerplate code (this can be written inline)
2) It's possible to provide multiple constructors for different packet variations to reduce the chance of errors.
3) It makes things catch fire on updates in ways that static analysers can understand.
This prevents plugins sending wrong packets at the compiler level (or would, if we had a compiler). It's more robust than a getter for client/server and throwing an exception since a static analysis tool can detect faults created by sending wrong packets from the server. This is also used to deny service to dodgy clients which send wrong packets to the server to attack it.
Something as simple as forgetting the reset() when encoding would cause lots of problems which could go unnoticed. This should be fully backwards-compatible but needs more tests.
