fix: normalize inverse-vol weights in float64

[Pigbibi]

Problem

Downcast research inputs can leave vol20 as float32. Inverse-vol normalization then rounds a fully invested target vector above the strict sum <= 1 + 1e-12 guard (for example, 1.0000001192092896), causing the producer export to fail closed.

Change

Convert vol20 to float64 at the inverse-vol weight calculation boundary before normalization. Existing finite, non-negative, no-leverage validation and missing-volatility filtering remain unchanged.

Validation

• Old implementation: characterization test reproduced the strict guard failure through clean_numeric_frame -> select_portfolio -> run_single_backtest.
• python -m pytest -q tests/test_backtest_accounting.py: 29 passed, 20 subtests passed.
• ruff check src/portfolio.py tests/test_backtest_accounting.py: passed.
• git diff --check: passed.

The real producer recovery run 34753427333 remains failed and is not represented as recovered by this local synthetic regression.