They don't store the parsed representation of the schema, since that would lock the table to a specific version. They store the create command itself, and each time you open a DB the schema is generated from that. Allows for flexible upgrades to the schema system without the need for migrations, and lets you use the same DB file with multiple sqlite versions.
The downside is that alter table commands are just edits to the create table string, which when accounting for all the different versions is fairly difficult and risky, so alter operations are limited.
SQLite is both.
Setting a compiler flag is reasonable for the common case where it is included as a header file in a C/C++ application.
It is certainly inconvenient when you design with that feature in mind, but can't rely on it being part of the standard binary distributions.
It doesn't really jibe all that well with things like Python though, where you typically don't do that. But it's certainly not a "foreign concept" or "almost meaningless".
For e.g. Python you can still just set the desired parameters when connecting, which should usually be handled by the database connector/driver for you.