Aggregate at an explicit grain

State the output grain before aggregating: one row per customer key in this exercise. Then distinguish how many observations exist from how many values are known. Group size includes rows whose amount is missing; count includes only known amounts.

A group with two missing amounts is not a group totaling zero. Use min_count=1 to preserve that distinction. Also decide what happens to missing group keys and arithmetic overflow; defaults should not silently discard or wrap evidence.

Cast cents to object/Python integers for summation before converting the reviewed result back to nullable Int64. This avoids a signed64-bit accumulator silently wrapping. A group whose exact total is outside the chosen output range is rejected explicitly.

dropna=False retains an unknown-key group in this analytical exercise. The later merge contract rejects unknown keys when establishing relationships. Grouping policy and join policy serve different purposes and should not be conflated.

import pandas as pd
frame=pd.DataFrame({"key":["zero","missing","missing"],"cents":pd.array([0,None,None],dtype="Int64")})
groups=frame.groupby("key",dropna=False)
print(groups.size().to_dict())
print(groups["cents"].count().to_dict())
print(groups["cents"].sum(min_count=1).isna().to_dict())

The output exposes rows, known_cents and total_cents separately. A zero total with one known observation differs from an unknown total with two missing observations. Keep both the count semantics and nullable result dtype visible in the report.

Aggregating at the wrong grain or dropping unknown keys can change later reconciliation. Test signed values, empty input, all-missing groups, zero groups and an overflow counterexample instead of only a convenient positive dataset.

import pandas as pd
values=pd.Series(pd.array([2**63-1,1],dtype="Int64"))
exact=values.astype(object).sum()
print(exact)
print("fits signed64",-(2**63)<=exact<=2**63-1)

Implement the rollup with count and missing semantics visible. Reject overflow in this bounded schema. Next enrich rows through a validated relationship.