Fix bug in PassThroughPropagator

[NthPortal]

Fix incorrect/invalid optimisation in PassThroughPropagator in which extracting values and then extracting nothing still left the values that were extracted the first time.

[iRevive]

Hm, OpenTelemetry Java keeps original context when no fields were extracted.

[iRevive]

I tried to replicate this test with oteljava PassThroughPropagator (the test fails):

test("does not inject fields after extracting none") {
  import scala.jdk.CollectionConverters._
  
  val getter = new TextMapGetter[Map[String, String]] {
    def keys(carrier: Map[String, String]): java.lang.Iterable[String] = carrier.keys.toList.asJava
    def get(carrier: Map[String, String], key: String): String = carrier.get(key).orNull
  }
  
  val setter = new TextMapSetter[scala.collection.mutable.Map[String, String]] {
    def set(carrier: scala.collection.mutable.Map[String, String], key: String, value: String): Unit =
      carrier.update(key, value)
  }
  
  val propagator = PassThroughPropagator.create("foo", "bar")
  val extracted1 = propagator.extract(Context.root(), Map("foo" -> "0"), getter)
  val extracted2 = propagator.extract(extracted1, Map.empty[String, String], getter)
  
  println(extracted1)
  println(extracted2)
  
  val map = scala.collection.mutable.Map.empty[String, String]
  propagator.inject(extracted2, map, setter)
  
  println(map)
  assert(map.isEmpty)
}

The output:

{passthroughpropagator-keyvalues=[foo, 0]}
{passthroughpropagator-keyvalues=[foo, 0]}
HashMap(foo -> 0)

[NthPortal]

that feels very wrong, I don't like that at all. the issue is, if you do

  val extracted2 = propagator.extract(extracted1, Map("bar" -> "1"), getter)

instead, it will remove "foo" -> "0" from the context, which seems inconsistent

[iRevive]

My expectation from the PassThroughPropagator is that it would try to extract all fields. Let's say the propagator is PassThroughPropagator("foo", "bar"), and I expect both foo and bar to be extracted from the carrier.

And a few scenarios and behaviors that I expect:

1. carrier = Map("foo" -> "1", "bar" -> "2") - the upstream propagated two headers -> List("foo" -> "1", "bar" -> "2") should be stored in the context
2. carrier = Map("foo" -> "1") - the upstream propagated only one header -> List("foo" -> "1") should be stored in the context
3. carrier = Map() - the upstream doesn't propagate any header -> nothing should be stored in the context

The previously stored fields should be overwritten in the context.

---

In some sense, it would be easier if the PassThroughtPropagator stored values as List[String -> Option[String]].
Then, the behavior would be more intuitive.

For carrier = Map("foo" -> "1") it would be List("foo" -> Some("1"), "bar" -> None).

[NthPortal]

after chatting with @rossabaker and looking at the spec I think I agree with you, though the spec is not worded super clearly