fix(torghut): block capital-unsafe replay feedback

[gregkonush]

Summary

• Blocks replay feedback that already violated realized capital constraints from being selected again as positive repair candidates.
• Propagates realized capital breaches through exact spec, signature, shape, and risk-profile feedback with hard block selection reasons.
• Keeps small down-day feedback eligible when drawdown and worst-day loss remain within the profit-target oracle policy, while treating excessive drawdown as a hard feedback blocker.

Related Issues

None

Testing

• uv run --frozen ruff format scripts/run_whitepaper_autoresearch_profit_target.py tests/test_run_whitepaper_autoresearch_profit_target.py
• uv run --frozen ruff check scripts/run_whitepaper_autoresearch_profit_target.py tests/test_run_whitepaper_autoresearch_profit_target.py
• uv run --frozen pytest tests/test_run_whitepaper_autoresearch_profit_target.py tests/test_portfolio_optimizer.py tests/test_candidate_specs.py tests/test_profit_target_oracle.py -q
• uv sync --frozen --extra dev
• uv run --frozen pyright --project pyrightconfig.json
• uv run --frozen pyright --project pyrightconfig.alpha.json
• uv run --frozen pyright --project pyrightconfig.scripts.json

Breaking Changes

None

Checklist

• Testing section documents the exact validation performed (or N/A with justification).
• Screenshots and Breaking Changes sections are handled appropriately (removed or filled in).
• Documentation, release notes, and follow-ups are updated or tracked.