[Bug] pass@k and avg@k collapse to attempt-1 accuracy in micro-weighted aggregation

I noticed that pass@k and avg@k are reported incorrectly when the aggregation mode is micro_weighted.

The metric helpers first compute the correct overall pass@k and avg@k values. However, aggregate_all_metrics passes the same per_cat_counts to accuracy, pass@k, and avg@k. In micro-weighted mode, aggregate_with_policy ignores the metric-specific per-category values and always calculates correct / total from these counts, which represent attempt-1 accuracy.

Here is a minimal example:

from types import SimpleNamespace

from agentcompass.runtime.metrics.helpers import aggregate_binary_metrics

results = [
    {
        "task_id": "t1",
        "category": "coding",
        "attempts": {
            "1": {"correct": False},
            "2": {"correct": True},
        },
    },
    {
        "task_id": "t2",
        "category": "coding",
        "attempts": {
            "1": {"correct": False},
            "2": {"correct": True},
        },
    },
]

print(aggregate_binary_metrics(results).metrics)
print(
    aggregate_binary_metrics(
        results,
        SimpleNamespace(avgk=False),
    ).metrics
)

The expected values are:

accuracy = 0.0
avg_at_k = 0.5
pass_at_k = 1.0

The reported values are:

accuracy = 0.0
avg_at_k = 0.0
pass_at_k = 0.0

I reproduced this on revision 16cb375779271d4524cde785d1656e1fc4528fac with Python 3.11.9. No model or external API is required.

A possible fix would be to preserve the correctly computed overall values in micro-weighted mode, or provide separate numerators and denominators for each metric instead of reusing the attempt-1 accuracy counts.

I'd be happy to work on the fix and add focused regression coverage for avg@2 and pass@2. Please let me know if this approach looks reasonable.

Start in agentcompass.runtime.metrics.helpers at aggregate_binary_metrics, then trace aggregate_all_metrics and aggregate_with_policy to see how per-category counts are reused. Reproduce the supplied two-result example and add focused regression coverage for avg@2 and pass@2. Done means micro_weighted aggregation reports accuracy 0.0, avg_at_k 0.5, and pass_at_k 1.0.