fix: make share, publish, et al. operator calls referentially transparent

[cartant]

Description:

This PR incorporates the fixes and tests in #5585 and also fixes and tests the share, shareReplay, publishBehavior and publishLast operators.

The problem that is fixed can be described as follows:

• Calls to operators - like share - must be referentially transparent. Otherwise, those operators cannot be used in calls to the static pipe function (to create partial pipelines).
• It's the calls to the operator functions - the functions that are returned from operator calls - that do not have to be referentially transparent (and can therefore implement the publishing and sharing). Those calls are made when the complete pipeline is composed via a call to a source observable's pipe method.

IMO, the changes are a bug fix, rather than a breaking change. The behaviour is changed. However, the behaviour prior to this change was not sensible and I cannot see how anyone could possibly be relying upon it. Any callers using a partial pipeline composed using the static pipe function and one of the affected operators would see wildly different behaviour - subscriptions to completely unexpected sources - depending upon the order in which subscriptions were made.

Related issue (if exists): #5411 #5585

[benlesh]

I'm generally in favor of this change. It seems like a bug that may have been introduced at some point. Can we prove that out? Is there a version of RxJS 5 or 6 where these new tests would pass?

[cartant]

It seems like a bug that may have been introduced at some point

Looking at the operators in HEAD of the 5.x branch:

• share is referentially transparent
• publish is not referentially transparent
• publishBehavior is referentially transparent
• publishLast is referentially transparent
• publishReplay is not referentially transparent

And it's the same with HEAD of the 6.x branch.

Going back further and looking at the first 'lettable' versions of the operators:

• share is referentially transparent
• publish is not referentially transparent
• publishBehavior is referentially transparent
• publishLast is referentially transparent
• publishReplay is referentially transparent

[benlesh]

So this is a regression in publishReplay introduced somewhere, and arguably a bug that has existed in publish since 5.x.

[benlesh]

I guess the workaround for this if you really WANTED this behavior in, say, publish would be:

const olderPublish = () => {
  const subject = new Subject();
  return connect({
    connector: () => subject
  })
}

Perhaps we should document that? Or maybe the comment here will be enough. I doubt it will be an issue, TBH.