Make the pii occurrence pagination test independent of numeric labels

[Travmatth]

Summary

test_occurences_pagination asserted that 12 random "The number is N" values all come back as NUMERIC_VALUE. With Solar's new English NER model (upbeat-leaf-281, TonicAI/solar#3838), some bare 3-digit numbers are labeled CREDIT_CARD:

• The model's credit_card_number tag scores 0.51–0.69.
• That beats the 0.5 regex NUMERIC_VALUE when the two overlap.

So the NUMERIC_VALUE count varied with the random numbers (5, 8 or 10 of 12) even though pagination worked. Reproduced on the #3838 PR app: on one seed, 7 of 12 numbers came back CREDIT_CARD, and the paginated and unpaginated fetches agreed.

The test now:

• uses 30 distinct numbers;
• compares paginated (limit=5) results to a single unpaginated fetch across all PII types, and for the mixed PiiType/str request;
• checks every number was found under some label.

It no longer depends on which numeric label the model picks.

Test plan

• Passed twice against the #3838 PR app running upbeat-leaf-281 (further runs were blocked by a local Border0 DNS outage)
• Solar's Python SDK job on TonicAI/solar#3838 picks this up from the default branch once merged

🤖 Generated with Claude Code