The launch check is a diff against a recorded baseline, not a health check

A health check asks "does it answer?" and the platform would have passed one all week. The thing that could break the demo is subtler: a DCID quietly withdrawn or redefined, which returns an empty result — the same shape as a country that does not report. So probe/launch_diff.py records a full snapshot (every crosswalk variable, every series value by value, every peer comparator, the enumerated corpus, the tool list) and compares the next run against it field by field.

Two consequences worth keeping. The baseline is only updated deliberately (--set-baseline), because a checker that re-baselines on every run reports "no change" forever. And the diff has its own self-test (--self-test), which injects each drift class into a copy of the baseline and asserts it is reported: a diff that returns "no change" and a diff that cannot see change look identical from the outside.