Reconcile noisy annotator labels

Question

Three annotators label each item and report an integer confidence. Produce the final label by weighted vote and count the items where every annotator agreed.

```python
def solve(raw: str) -> str: ...
```

- The first line is the item count `n`. Each of the next `n` lines has exactly three `<label>:<confidence>` pairs, where the label is `A` or `B` and the confidence is a nonnegative integer.
- Sum the confidence per label. Return `A` when its total is higher, `B` when its total is higher, and `ABSTAIN` on an exact tie.
- After the per-item labels, return one final line `agreed=<count>`, where an item is agreed only when all three raw labels match.
- Return one label per item in input order, then the agreement line. Do not append a trailing newline.