Every distinct production error gets its own ticket

The incident list is supposed to hold one entry per distinct error a portal reports. It was holding one entry per component: the identity of an incident was built from where the fault was raised and what type it was, and nothing else — so thirteen NodeTypes failing to compile, each with its own compiler errors and its own fix, arrived at the same stack frame with the same exception type and became a single ticket. None of the thirteen was reported.

An incident is now identified by the fault's own words as well: the exception's message, with everything that varies between occurrences masked out — node paths, ids, counts, elapsed times, and any name the message itself uses as a path. A thousand repeats of one error are still one entry with a rising count; two genuinely different errors are two entries with names you can tell apart. When one log source produces so many different shapes that the masking is clearly missing something, they fold into a single entry that says how many it stands for, rather than flooding the list.

A window the watcher could not read completely now reports that as a finding of its own, as does a stretch skipped after a long outage. Silence from the log store is no longer mistaken for quiet.