fix: preserve walkforward warmup counts

[Pigbibi]

Problem

When walk-forward predictions exist, joining the aggregated prediction frame leaves prediction_window_count as NaN for the pre-OOS prefix. build_final_scores correctly keeps those rows out of final scoring, but run_single_backtest requires prediction_window_count == 0 to recognize model warmup. The prefix is therefore misclassified as incomplete eligible scores before the first OOS window.

Change

Normalize only unmatched prediction_window_count values to integer 0 after the prediction join. Scores remain untouched; positive prediction counts with missing scores still fail the existing strict guard.

Validation

• RED: an actual run_walkforward_scoring -> build_final_scores -> run_backtest_suite fake-panel path reproduced ValueError: Backtest has incomplete eligible scores with pre-OOS counts as NaN.
• GREEN: the same path keeps pre-OOS final_score as NaN, sets counts to 0, and completes the backtest suite.
• python3 -m unittest discover -s tests: 131 tests OK.
• python3 -m pytest -q tests/test_walkforward_validation.py tests/test_backtest_accounting.py: 39 passed, 40 subtests passed.
• python3 -m ruff check . and git diff --check: passed.

No producer/consumer workflow, provider, VPS, or trading operation was run.