The checks now notice when they start seeing less

A gate can be green for two reasons that look exactly alike from outside: the change was safe, or the thing that would have failed was never in the list.

The platform's cross-repo checks — the ones that stop a change removing a public type here from breaking a repository that uses it — all ask their question of one index: every public type and public member declared under src/. The index is built by reading the source files. So the checks are only ever as good as that reading, and a reading can go wrong in a way that produces no error, no warning and a perfectly plausible number.

What happened

Three invisible bytes at the start of a file — a byte-order mark, which many editors add without being asked — put the very first character out of reach of a pattern anchored to the start of a line. 310 of the platform's 1 271 compiled source files carry one.

The effect was not that the index failed. It was that the index held 1 955 types where the truth was 1 971: sixteen public types were in it nowhere at all, and a hundred and twenty-four more were filed under the wrong name. The checks read that number, found it healthy, and answered confidently.

Nothing slipped through — that was measured rather than hoped for, two different ways, across every one of the 754 changes made while the gap was open. But it was a reprieve, not a defence: the checks had simply not yet been asked the question they would have got wrong.

What changed

The reading now has to prove it did not get worse.

Whenever a change touches the code that builds the index, the platform runs the previous version of that code and the new version over the identical set of files, and compares what each one found. Because both are reading exactly the same files, any difference is the reading — not the codebase.

• Seeing fewer things than before stops the change, and says which counter dropped and by how much. If the old number really was too high, that gets written down as a sentence in the change itself, per counter, with a reason.
• Seeing more is a fix, and passes. The repair that started all this measures +16 types and +111 members over an identical set of files — which is the entire point of it.
• Seeing the same is printed with both numbers, so a pass says something instead of nothing.

There was already a check that the index is not empty. That one catches "nothing was read". This one catches "less was read than last time", which is the failure that actually happened, and which no threshold on a single number can see: the gap was under one percent of the total.