Fix 10 bugs from QA audit: drawdown, PDT, Sharpe, reviewer alignment

[mikejamescalvert]

Summary

Deep dive QA audit across all 38 source files found ~80 issues. This PR fixes the 10 highest-impact ones (quick wins + strategy alignment).

Bug Fixes (8)

• Drawdown returns negative values — EquityTracker.current_drawdown() went negative when equity exceeded peak, confusing the circuit breaker. Now clamped to 0.
• Peak equity lost on file corruption — If peak_equity.json was missing/corrupt, peak reset to 0 forever. Now recovers from equity history.
• REDUCE_SIZE multiplier was 0.75, docstring said 0.50 — Changed to 0.50 (more conservative during drawdown, which is correct for a small account)
• Sharpe ratio inflated ~5% — Was using population std (÷N) instead of sample std (÷N-1)
• Performance report used 30-day window — Statistically meaningless. Changed to 90 days.
• PDT protection silently broken — str(o.side) might produce "OrderSide.BUY" not "buy". Now uses .value for reliable enum extraction.
• Cooldown timezone bug — Naive vs aware datetime comparison could raise TypeError, silently disabling anti-churn protection.
• Cooldown dict grows forever — Added pruning of expired entries (>72h) on load.

Strategy Alignment (3)

• Portfolio reviewer gave defensive advice — Prompt emphasized "diversification" and "concentration risk" for a momentum strategy that WANTS concentration. Rewritten to focus on capital deployment, momentum alignment, and speed to alpha.
• Review issue dedup too aggressive — Category-based dedup blocked ALL new suggestions if any issue in that category existed (e.g., one "risk" issue = no more risk suggestions ever). Removed. Title-based fuzzy matching still prevents true duplicates.
• Watchlist reviewer could drop held positions — AI was only asked to keep them (prompt instruction). Now enforced in code — held positions are always merged back into the watchlist after AI review.

Test plan

• All 270 tests pass
• Ruff lint clean
• mypy strict clean (38 source files)
• Updated 3 tests to match new behavior (drawdown clamp, multiplier change)
• Verify next portfolio review creates actionable momentum-aligned suggestions
• Verify PDT protection works correctly (check logs for order_blocked_pdt)

🤖 Generated with Claude Code