Proposal #96: Stabilize in-epoch invalid SPRT on small samples

In-epoch SPRT on invalid inferences is not stable on a small sample. A few invalids can trip it before the epoch has enough data. This proposal raises invalidation h threshold to 40 and sets bad participant invalidation rate to 18e-2. At large sample sizes the threshold is about 10% invalid, so participants below that are not punished. Higher rates still trigger, 18% in about 360 samples and 12% in about 1650. Small samples cannot trigger it. At least 32 flagged invalids are required, and at 500 samples the observed rate must exceed 16%. Downtime SPRT is unused and is turned off. Epoch-end stats tests still work.