Flaky integration test: SessionConfigLive.StaticExchangeDeliversBothPayloads fails ~2 in 10 runs #62

Observed while working on #58; the failure predates that branch.

Symptom

SessionConfigLive.StaticExchangeDeliversBothPayloads (Tests/Integration/SessionConfigLiveTests.cpp) fails intermittently:

Tests/Integration/SessionConfigLiveTests.cpp:280: Failure
Expected: (incoming) != (nullptr), actual: NULL vs (nullptr)
[  FAILED  ] SessionConfigLive.StaticExchangeDeliversBothPayloads (15245 ms)

The failing run takes ~15 s, which is the test's wait deadline, so one side never observes the expected packet rather than observing a wrong one.

Rate

• On master itself (commit 29b35a10, master's own test binary, macOS arm64): 2 failures in 10 consecutive runs with --gtest_repeat=10.
• On the Reduce core surface: dead platforms, extras library, DS_* deprecations, snake_case files (#50) #58 branch: 1 in 5.
• CI's ctest --repeat until-pass:3 on the integration label currently absorbs it, which is exactly the kind of masking that retry is not meant for (CLAUDE.md: retries absorb timing misses, never real bugs). We do not yet know which this is.

To do

1. Reproduce with ./build/Tests/IntegrationTests --gtest_filter='SessionConfigLive.StaticExchangeDeliversBothPayloads' --gtest_repeat=50 --gtest_break_on_failure and capture which side is missing which message (add temporary logging of every packet id received by both peers until it fails).
2. Decide between a test defect (both sides not pumped until each has observed the event, a wait that is too short, a fixed port collision with another suite) and a library defect in the session-config exchange (a withheld connection packet never produced when the session handshake and the connection request interleave in a particular order).
3. Fix at the right layer; if it is the library, add a hermetic reproduction to the unit suite using the two-layer fake-socket harness rather than relying on loopback timing.

Notes

• A superficially similar Linux-only timeout in DirectoryDeltaTransferStreaming during the same period turned out to be a real bug (an unbounded directory walk), not a flake; it was fixed in Reduce core surface: dead platforms, extras library, DS_* deprecations, snake_case files (#50) #58. Treat this one with the same suspicion until the cause is known.