Summary

Adds a new opt-in turnover_aware portfolio optimizer under backtest/optimizers/. It maximizes a mean-variance utility minus an L1 penalty on the change in weights relative to the previously held weights, solved with SLSQP on the long-only simplex.

This is the first step of Portfolio Studio (#456), scoped per the maintainer guidance there: turnover-aware optimization first, existing four schemes untouched, config-driven parameters, changes confined to backtest/.

Why

The engine already loads optimizers dynamically via config.optimizer → backtest.optimizers.<name>.optimize(ret, pos, dates, **optimizer_params) (backtest/engines/base.py), with four schemes today (mean-variance, risk parity, equal-volatility, max-diversification). None of them account for turnover cost: because weights are recomputed on a rolling window each rebalance, an optimizer can churn positions every bar, so an optimized backtest can overstate net returns once real transaction costs apply. This optimizer lets the objective trade off expected risk-adjusted return against turnover, and reports realized turnover so cost-adjusted performance is visible.

Changes

• backtest/optimizers/turnover_aware.py: new TurnoverAwareOptimizer(BaseOptimizer) + module-level optimize(). Objective min −wᵀμ + λ·wᵀΣw + γ·‖w − w_prev‖₁, long-only simplex via SLSQP, warm-started from prior weights, falling back to equal weight on non-convergence / NaN covariance. Prior weights are tracked over the full universe (entries and exits both count as turnover). Realized per-rebalance turnover is accumulated on the instance.
• backtest/optimizers/__init__.py: document the new scheme.
• tests/test_turnover_aware_optimizer.py: 12 tests.

Notes:

• Opt-in only; default backtest behavior is unchanged (optimizer off → 1/N).
• With turnover_penalty=0 the objective is path-independent and equals the un-penalized mean-variance-utility solution (verified against a brute-force simplex search).
• The penalty is scale-sensitive relative to the return term; this is documented in the module docstring so callers can tune it to their data frequency.
• No new dependencies (reuses numpy / scipy.optimize, already used by mean_variance.py).

Test Plan

• ruff check — clean
• pytest tests/test_turnover_aware_optimizer.py — 12 passed
• pytest full suite (excl e2e) — 4895 passed, 9 skipped (2 pre-existing env failures unrelated to this change: test_factor_gate_fund_prefix, test_factor_operators::test_bottleneck_available, both reproduce on clean main)

Test coverage: zero-penalty path-independence, higher penalty lowers realized turnover, turnover monotone non-increasing across penalty levels, long-only simplex feasibility, input sign preservation, short-window / NaN / all-NaN-column graceful handling, single-asset passthrough, realized-turnover reporting.

Checklist

• No changes to protected areas without prior discussion (direction approved in [Feature] Portfolio Studio: turnover-aware optimizer + constraints (planning) #456)
• No hardcoded values (all parameters config-driven with defaults)
• Code follows CONTRIBUTING.md guidelines (type hints, Google-style docstrings, DCO sign-off)
• Documentation updated (package docstring lists the new scheme)

Refs #456