By moving the duplicate file removal to the manipulate_files coroutine, we
ensure that all previous albums are fully moved/copied before trying to delete
their duplicate files.
This new design lets us provide better common functionality. The `get()`
method is a particularly good example: we can now avoid the try/except dance
that was previously necessary to get the only result for a single-result
query. This also lets us continue using `len()` on the result of
`album.items()`, which previously returned a materialized list.
A second base class, LibModel, maintains a reference to the Library and should
take care of database-related tasks like load and store. This is the beginning
of the end of the terrible incongruity between Item and Album objects (only
the latter had a library reference). More refactoring to come.
One large side effect: Album objects no longer automatically store
modifications. You have to call album.store(). Several places in the code
assume otherwise; they need cleaning up.
ResultIterator is now polymorphic (it takes a type parameter, which must be a
subclass of LibModel).
I added this back when I thought I would implement a different subclass that
represented iPods. That means I wrote this code when I had a lime iPod mini. I
am old.
In preparation for enabling queries over flexattrs, this is a new path that
lets queries avoid generating SQLite expressions altogether. Any query that
can be completely evaluated in SQLite will be, but when it can't, we now fall
back to running the entire query in Python by selecting everything from the
database and running the `match` predicate.
To begin with, this mechanism replaces RegisteredFieldQueries, which
previously used Python callbacks for evaluation. Now they just indicate that
they're slow queries and the query system falls back automatically.
This has the great upside that it lets use implement arbitrarily complex
queries without shoehorning everything into SQLite when that (a) is way too
complicated and (b) doesn't buy us much performance anyway. The obvious
drawback is that any code dealing with queries now has to handle two cases
(slow and fast).
In the future, we could optimize this further by combing fast and slow query
styles. For example, if you want to match with a substring *and* a regular
expression, we can do a first pass in SQLite and apply the regex predicate on
the results. Avoided for now because premature optimization, etc., etc.
Next step: implement flexattr matches as slow queries.
Namespaces were a worthy idea, but they added a lot of complexity to both the
library code itself and every client of the flexattrs interfaces. Getting rid
of them, and having one flat namespace of both traditional fields and
flexattrs, has one huge benefit: we can "promote" flexattrs to real attributes
(and vice versa) without code changes in every client.
This frees us to have a somewhat less efficient implementation of flexattrs
because we have a smooth upgrade path for making attributes more efficient via
promotion.
