Sometimes a single Roslyn Emit throws and, from that moment, every later compile in the same process fails identically. Parsing and binding keep working perfectly; only the emit is dead. One such event surfaces as up to a dozen unrelated test failures, and each time it has cost a fresh — and always identical — misdiagnosis of whatever change happened to be under test.

The compile pipeline already detects this and logs one loud line saying so: "every later compile in this process will fail the same way — attribute the failures that follow to this line, not to the change under test". But it logged that line without the exception object, and the complete log a CI run retains only records entries that carry one. So the loudest statement the pipeline produces was also its least durable: absent from that log by construction, and absent from the per-test output whenever the test that happened to be running had passed. Measured on two real occurrences: 25 copies of the line in the per-test output, zero in the retained log, both times.

Passing the exception is the whole fix — the same thing the neighbouring call already did.

The self-diagnosis that runs at the first failed emit also grew a read it was missing. It checked that a nested type's container reads back correctly, but never that a top-level type's container reads back as absent — which is precisely the check the metadata writer makes before the call that crashes. A process broken in that direction would have been reported healthy. It now reads both directions, and says which one it read.