fix(backtest): keep excess_return consistent with corrected benchmark_return

[Shizoqua]

Summary

• BaseEngine.run_backtest() corrects the displayed benchmark_return once an external benchmark's price series contains a non-positive-prior-price bar, but never recomputes excess_return from that same corrected value
• Recomputes excess_return right where benchmark_return is already corrected, so the two fields remain consistent for this corrected benchmark path

Why

Issue #872 replaced the fragile pct_change-based return calculation with a safe price relative (buy_and_hold_return) for the displayed benchmark_return, after finding it could report numbers like +39,560% where the true return was -42.7%. calc_metrics() has its own internal re-derivation of the same total, still using the exact (1 + bench_ret).prod() - 1 pattern #872 found unsafe. base.py corrects the displayed field but never re-derives excess_return from the corrected value, so the metrics dict can become self-contradictory: benchmark_return reads correctly while excess_return implies a different number, both in the same JSON output the backtest tool returns to the LLM.

I audited every benchmark-derived metric in calc_metrics() (benchmark_return, excess_return, information_ratio, tracking_error, benchmark_beta). Only excess_return is affected. The other three are computed from the per-bar bench_ret series directly, which is already the corrected, #872-safe series by the time it reaches calc_metrics().

Changes

• agent/backtest/engines/base.py: after the existing benchmark_metadata override, excess_return is recomputed as total_return minus the corrected benchmark_return, but only when an external benchmark was actually resolved. The internal-basket-benchmark path (no config["benchmark"]) is untouched.
• agent/tests/test_benchmark_excess_return_consistency.py: two new regression tests. One runs the real production path (ChinaAEngine.run_backtest, real calc_metrics(), only the network-dependent resolve_benchmark() fetch is stubbed) with a benchmark price series crossing a non-positive-prior-price bar, and asserts total_return minus excess_return equals benchmark_return. The other is a control confirming the fix does not touch excess_return when no external benchmark is configured.

Test Plan

• Confirmed the load-bearing regression test fails on the pre-fix implementation with a real numeric assertion failure (excess_return read 1.0 instead of 0.5), and passes with the fix
• pytest tests/ -k "engine or backtest or benchmark or metrics" -q --ignore=tests/e2e_backtest (813 passed, 6 skipped, pre-existing)
• black --check on both changed files: pre-existing, unrelated formatting issues in base.py confirmed via diff-isolation against the unpatched file (identical content at shifted line numbers); the new test file is clean
• ruff check on both changed files: all checks passed

Checklist

• No changes to protected areas (src/agent/, src/session/, src/providers/): this only touches agent/backtest/
• No hardcoded values
• Code follows CONTRIBUTING.md guidelines
• Documentation updated (n/a: internal correctness fix, no public interface change)
• DCO signed-off