bug: exec_accuracy under-scores company_facts when filter returns multiple rows

[SebAustin]

---

title: "bug: exec_accuracy under-scores company_facts when filter returns multiple rows"
labels: [bug, eval, w2]
assignees: [SebAustin]

Problem

Live edgar-msft-revenues-fy2025 can return several FY2025 rows for
RevenueFromContractWithCustomerExcludingAssessedTax. Deterministic
exec_accuracy requires an exact list length match against expected, so
the case scores 0.0 even when the judge confirms the correct fact is
present.

Acceptance criteria

• exec_accuracy (or a company-facts-specific helper) matches when
  expected.xbrl_facts is a subset of actual.xbrl_facts by concept,
  fiscal year, and value tolerance.
• Unit tests in tests/unit/evals/test_metrics.py cover multi-row
  company-facts payloads.
• Live --full run reports mean_exec_accuracy >= 0.95 without
  lowering judge scores.

References

• src/mcp_financial_data/evals/metrics.py
• src/mcp_financial_data/evals/dispatch.py::_filter_xbrl_facts_for_eval