feat: portfolio signal contract + by-ref loader + sizing

[ArthurBernard]

Summary

• application.portfolio — the multi-asset analogue of the single-instrument signal:
  • PortfolioSignalFn = (asof_ms, {Symbol: frame}) -> {Symbol: weight} (weight = signed fraction of capital);
  • PortfolioStrategy (frozen: universe + signal + capital + optional gross_cap);
  • weights_to_signals(weights, *, prices, capital, asof_ms) — pure sizer, qty = weight × capital / price → Signal.target_qty (exact Decimal, never float);
  • load_portfolio_signal(ref) — safe module:function loader (importlib+getattr, no loose-file exec).
• Reusable fake PortfolioSignalFn fixture (tests/fixtures/fake_book.py) for downstream leaves.

Why

LS1 (first validated multi-asset strategy) is a vector of target weights over ~10 coins, gross-capped 2×; the single-instrument (bars) -> Signal can't express it, and a per-coin |w|>1 (leverage) would be rejected by a bounded [-1,1] exposure signal. An explicit-quantity signal carries its own scale, so the runner diffs it against live positions with no reference_qty. The loader stays by-reference so the engine is generic (LS1 lives in config). See ADR.

Changes

• trading_bot/application/portfolio.py (new), application/__init__.py (exports).
• trading_bot/tests/fixtures/{__init__,fake_book}.py (new), tests/application/test_portfolio_signal.py (new, 20 tests).

Changelog

Added under [Unreleased] (0.3.0): application.portfolio weight-vector contract + sizer + loader.

Test plan

• .venv/bin/python -m pytest — 637 passed, 7 deselected
• .venv/bin/ruff check trading_bot/ — clean
• .venv/bin/mypy trading_bot/ — clean
• CI green

Leaf 01 of portfolio-strategy (doc/dev/plans/portfolio-strategy/). Leaves 02 (feed) and 03 (runner) depend on this.