Two robustness fixes after the census flaked under full parallel probe
load (not a product regression — the failing diff was '# transaction' vs
'# transactions', pure pluralization):
- normalize collapses count-driven plurals so a data-count of 1 vs N no
longer churns the snapshot (the census tracks controls, not the data in
their labels)
- settle waits for the nav shell + 3 stable count samples (was 2), and the
project retries 2x — deterministic snapshots mean a real control change
still fails every retry while a load-flake passes
Re-baselined from a clean seed. Probe 33/33 stable across runs.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
1f6073823b