[SimonCropp]

ListenerTableTests.AListenerIsListedOnEveryPlatform failed on the windows job of VerifyTests/DiffEngine#1015, on net48, which touched nothing near it. After stopping its listener it still found the port listed. Two separate problems can make that happen.

1. Windows hands out a port again as soon as it's freed. The tests beside this one bind port zero, so one of them can be given the port the moment it's let go. Locally, AListenerSaysWhoseItIs found its stopped listener's port in use by another listener in the same process within milliseconds.

• New HeldPort (in the tests): a port bound but not listening, so nothing else can be given it.
• HeldPort.AfterListening runs a test's question while its listener is up, stops the listener, and binds the port again before asking what the table says about it. If something got the port first, it starts again on a new port, up to ten times.
• It's used by the three tests that ask about a port after stopping its listener:
  • ListenerTableTests.AListenerIsListedOnEveryPlatform;
  • ListenerTableTests.AListenerSaysWhoseItIs;
  • TrayDetectorTests.AListenerOnThePiperPortIsATray.

2. The tests' replacements of ListenerTable.Lookup could leak. Lookup was a delegate. Each test replaced it with one that wrapped the one before, and put the old one back afterwards. When two tests overlapped, the second to finish restored the first one's wrapper, which then answered for its port for the rest of the run. ViewerProtocolTests.AConnectGivenUpOnIsObserved said a port it had already freed was listed, so whichever test got that port next was told so.

• ListenerTable.Answer(port, answer) replaces the delegate. It keeps one answer per port in a map, and disposing what it returns removes that answer. A second answer for a port that already has one throws, so two tests that think they hold the same port find out.

• ListenerTable.Listed is now internal, for a test answer that passes through to the real table.

• The callers moved to it:

  • TrayDetectorTests;
  • ViewerProtocolTests;
  • the Lookup helper in ViewerClientUnownedTests;
  • TrackerSnapshotTest's two tests.

• Ports are held now: a test answers only for a port it holds. The three that freed a port before answering for it use a HeldPort instead:

  • AConnectGivenUpOnIsObserved;
  • ATableThatCannotBeReadLeavesTheConnectToAnswer;
  • APortAConnectFoundEmptyIsNotAskedAboutAgainSoSoon.

  The last one lets its port go only to listen on it a moment later.

• New tests in ListenerTableTests: AnswersThatOverlapLeaveNothingBehind and APortIsAnsweredForOnce.

Tests. Locally on Windows, a Release build had no errors:

• DiffEngine.Tests ran twice. All 2,294 tests passed except 2 failures in the first run and 4 in the second. Every one stopped at an assertion on ListenerTable.Enumerated, which reads .NET's listener list. That list misses app listeners on this machine's Windows 10.0.26300 and fails the same way on main.
• DiffEngineTray.Core.Tests passed 321, with 4 skipped.
• With only the fix to Lookup in place, a full run failed AListenerSaysWhoseItIs on net48: "Expected to be null but found 11160", another listener in the same process on the freed port. That's what found the first problem. It didn't fail in either run after HeldPort.AfterListening.

Not seen working. I didn't reproduce the leaked replacement itself; the second problem is read from the code. AListenerOnThePiperPortIsATray runs only on Linux and macOS, so CI is the first run of its new form.