Adherence to Semantic Versioning is just The Right Thing To Do, but it turns out you have to be extra-careful when modifying public interfaces to maintain backwards compatibility. This is obvious on reflection, but I never thought about it beforehand. Thanks to Hari KT for pointing it out.

Why do you have to be extra-careful with interfaces and SemVer? To see it more clearly, let's use a concrete class as an example.

class Foo
    public function bar() { ... }
    public function baz() { ... }

If we remove a public method, that's clearly a BC break. If we add a non-optional parameter to an existing public method, that's also clearly a BC break. Those kinds of changes will break any code already using the Foo class; that code will have to be modified to accommodate the changes in Foo.

However, if we add a new public method to the concrete class, that is not a BC break. Likewise, changing the signature of an existing method to add an optional parameter is not a BC break either. (UPDATE: See note below.) Code using the Foo class does not have to change to accommodate the new method or the new optional parameter.

But what happens with an interface?

interface FooInterface
    public function bar();
    public function baz();

Removing a method, or adding a non-optional parameter to an existing method, is the same as with a concrete class: it's BC break.

However, unlike with a concrete class, adding methods is also a BC break. So is changing an existing method sigature to add an optional parameter. Existing code that implements FooInterface will no longer comply with the interface; that code will have to change to accommodate the new method or the new optional parameter.

Thus, interfaces are more susceptible to BC breaks than concrete classes. Once an interface is published as "stable", I don't see how it can be changed at all per the rules of SemVer (that is, unless you bump the major version). The only thing you can do to maintain BC is add an entirely new interface to the package while leaving the old one in place, perhaps even extending the old one if needed.

Again, it's obvious in hindsight, but I did not have the foresight to anticipate it. Perhaps this realization will save you some trouble in the future.

p.s. On further examination, the interface constraints apply to abstract classes, too. Changes to an abstract mean that child classes will have to be modified as well, otherwise they won't comply with the modified abstract.

UPDATE: Dave Marshall notes on Twitter that in the concrete class case, if you add an optional parameter to a method, and a user has extended and overridden that method, the extended class will break under PHP E_STRICT. He links to Symfony policy on the subject, which I will now have to read.

Are you stuck with a legacy PHP application? You should buy my book because it gives you a step-by-step guide to improving you codebase, all while keeping it running the whole time.