I think I understand switch-reference. It is where the verb is marked for whether or not the subject is the same as in the previous clause, right?
Except that for many or most languages with switch-reference marking, the reference clause is after the marked clause instead of before it. (The direction in which the switch-reference refers is statistically correlated with the "word"-order.)
Also, the reference clause may be the anchor clause (initial clause in some 'langs, final clause in others) of the clause-chain, rather than the immediately preceding or immediately following clause.
Also, sometimes the switch-reference marker specifies not only "same subject" and "different subject", but also or instead "same object" vs "different object". (Mostly even ergative switch-reference-marking languages track "same agent vs different agent", but not all do.)
And, sometimes it marks whether the marked clause's subject is the same as or different from some other core participant of the reference clause; that is some of its values may mean "marked clause's subject is referenced clause's direct object" or "marked clause's subject is referenced clause's indirect object", as well as "marked clause's subject is referenced clause's subject" and "marked clause's subject is not a core argument of referenced clause".
And sometimes the marker distinguishes between identity and proper containment. That is, in addition to a value meaning "marked clause's subject is referenced clause's subject", another value might mean "marked clause's subject is one of or some of referenced clause's subject" if the referenced clause's subject is non-singular, and another value might mean "marked clause's subject contains referenced clause's subject" if the marked clause's subject is non-singular, and another value would mean "marked clause's subject neither contains nor is contained in referenced clause's subject".
You could look up what I wrote about Adpihi's (my conlang's) switch-reference system
; it's quite typical except for one thing.
In Adpihi one value of the marker means both/either "marked clause's subject properly contains referenced clause's subject" and/or "marked clause's subject is properly contained in referenced clause's subject" and leaves the addressee to disambiguate based on gender and grammatical number; but another value means "marked clause's subject is exactly identical to referenced clause's subject".
But, in most natlangs with switch-reference marking that marks proper containment, one direction of proper containment is marked just like identity, and the addressee must disambiguate somehow; while the other direction of proper containment is marked differently.
How does it interact with relative clauses?
Subordinate clauses don't occur in languages with switch-reference marking (this sentence may be "almost true" instead of "true").
A clause is subordinate to another clause if both:
it is contained in it and plays a role in it (as if a noun or an adjective or an adverb) and;
it is dependent on it (its semantics depend on the semantics of the containing clause).
In clause-chaining switch-reference-marking languages a "subordinate" clause must be either the first or the last element of the matrix clause it's embedded in and dependent on; and there can be only one such "subordinate" clause. But the languages don't strongly distinguish between co-ordination and sub-ordination. You might call it "cosubordination".
Do I have any significant misconceptions?
I can't tell. AFAIK nothing you've said so far indicates any significant misconception.
BTW; if your switch-reference-marking system tracks two participants of the marked clause (like Adpihi's tracks both the marked clause's subject and its object), and mark their relationship possibly to more than one participant of the referenced clause, the clauses in a clause-chain will have to satisfy the "principle of disjoint reference". That is to say, if such a clause has two or more core participants (such as a subject and an object), then the individual or group referred to by one core participant (for instance the subject) cannot be the same as, nor be contained in, nor contain, nor overlap with, the individual or group referred to by any other core participant (such as the object).
Of course that will require that the language have good productive reflexive voice and reciprocal voice processes.
(I highly recommend this book.)http://people.wm.edu/~jbmart/papers/cr_switch_reference.pdfhttp://www.sil.org/acpub/repository/Ke-A_switch_reference_marker_Mankanya.pdfhttp://en.wikipedia.org/wiki/Koasati_languagehttp://www.wikinfo.org/index.php/Koasati_language
maybe; that link may be broken.http://ling.uni-konstanz.de/pages/proj/Sprachbau/introduction/examples.html
but I'm not sure how strong these statistical implications are.http://pilarvalenzuela.com/Shipibo_Language.html