How do you recommend or require in a RDFS or OWL ontology to use, say, dc:title to give a title to instances of a given class?

Here's my suggestion: Don't try to make such "recommendations" part of the model itself. What you are about here is a "soft" pragmatic aspect of your model ("one should use dc:title to enable search engines to get the most out of it"), but when you add axioms, then you make it into a "hard" semantic aspect. Instead, write up a "best practices" document, explaining what should be done, and maybe implement some lint-style checking tool to be used by others for checking that these best practices are being applied. If not, then this tool should give a warning, which is a lot better than having direct semantic consequences, such as an inconsistent knowledgebase, which may stop people from doing what they want to do, because they may have their good reasons not to follow your recommended best practices.

An alternative would be to introduce a special set of OWL annotations to indicate what should be done. OWL already has annotations for deprecation and for (in)compatibility between ontologies, which do not have a real effect on the model, and one could easily extend this set of annotations to whatever one wishes to have. One could and should do this outside the core OWL standardization process and may do it even outside W3C, but it would require some kind of standardization effort to make it into something that is widely used. So this would certainly be more of a future thing.

So the question is that you want to communicate to the modeler that they should use dc:title, and you want some machine operable way to detect when this convention is violated. Using RDFS/OWL you could set up a way to create an inconsistency when dc:title is missing. For example, infer all members of classes without a dc:title to be members of some "non-title" class and assert that all classes are owl:disjointWith that class. Any instance that is a member of the "non-title" class and one of the classes it is disjoint with will be flagged as an inconsistency.

There are two issues with that solution. First, computing complements in OWL is both troublesome (some may disagree with me on this) and computationally expensive. Second, reasoners tend to poorly support how inconsistencies are reported. At best you will get a list of model inconsistencies that include your convention and you get to sort out which ones you want to expose to the modeler creating the instance.

An alternative, which is more general than the notion of model consistency, is the SPIN notion of constraint checking. This defines a constraint check in a SPARQL query that can directly specify the semantics you need. For example, your rule would be:

# instance must have a dc:title property
ASK WHERE
{  FILTER NOT EXISTS  { ?this dc:title ?title .} .  
}

The rule is defined for a class and is applied to all members of the class, pre-binding ?this to each instance before running the rule. Any instance not having a dc:title is flagged as having a constraint violation. This could be displayed on the instance view of your editor. TopBraid Composer and the Web-based application builder TopBraid Ensemble both do this. I.e. when the user creates an instance and forgets dc:title, a warning appears in the view. In addition the comment line is used in the warning, giving you a way to communicate the rule's intent. (I work for TopQuadrant, so feel free to ask more questions here or on our forum.)

Either of these approaches would work, but I find the consistency check method to be more aligned with your needs.

I worry about the statement

I can make MyClass a subclass of the things which have a dc:Title - but not all instances of MyClass have a title

…as it seems to assume that instances inherit object attributes. This is simply not the case in RDFS/OWL. Unlike attributes in OO modeling, adding a property to a class definition in RDFS/OWL will have NO IMPLICATIONS on instances. Instances can have any eoprrties and the only way it is known that an instance is a member is through the type triple - {:thing rdf:type :someClass}. You may understand this already, but I'm not getting the gist of the solution you propose so I thought I'd throw this out as a possible misconception.

In Anzo, we do this via property restrictions, though I don't know how legit this is. So we'd do:

ex:myClass rdfs:subClassOf [
   a owl:Restriction ;
   owl:onProperty dc:title ;
   owl:minCardinality 0
]

...which, semantically speaking, really doesn't say anything at all (I think), but which our tools interpret as a hint to consider using this property (dc:title) in conjunction with this class (ex:myClass).