fix(models): Kelly sizing returns None when historical data is absent (#17)

[bradsmithmba]

Summary

ScoringEngine._calculate_kelly_sizing() was called with call-site defaults when a strategy had no historical performance data:

kelly_size = self._calculate_kelly_sizing(
    metrics.get('win_rate', 0.5),
    metrics.get('avg_win', 1.0),
    metrics.get('avg_loss', 1.0)
)

With those defaults, Kelly = (0.5*1 - 0.5) / 1 = 0. Every strategy lacking backtest data was sized at exactly 0 — indistinguishable from a genuine "no edge" result (Kelly is also 0 for a true 50/50, 1:1 bet) — and surfaced to the user as "Recommended position size: 0%", which reads as advice rather than missing data.

Closes #17.

Fix

_calculate_kelly_sizing(metrics) now takes the metrics dict and returns Optional[float]:

• None when win_rate, avg_win, avg_loss are not all present (no statistical basis to size).
• 0.0 when the data is present but there is genuinely no edge.
• a positive fraction otherwise (unchanged behavior).

ScoredStrategy.kelly_size becomes Optional[float]. recommendation_engine renders None as "insufficient historical data to size" instead of formatting it as 0.0%.

Tests

$ pytest tests/models/test_scoring_engine.py -q
11 passed

Two new tests pin the distinction:

• test_kelly_sizing_none_without_historical_data — missing inputs → None
• test_kelly_sizing_zero_for_no_edge — present inputs, no edge → 0.0

CI note

src/models/recommendation_engine.py currently cannot be imported because it imports the missing src.models.integrated_selector (#21), so its None-rendering branch can't be exercised in CI yet. The change there is a defensive guard for when #21 lands; the scoring-engine behavior and its tests run independently and pass now.

🤖 Generated with Claude Code

[bradsmithmba]

Dependency note: the scoring-engine tests in this PR pass on their own (scoring_engine only imports strategies.base). The one piece that can't be exercised in CI yet is the recommendation_engine None-rendering branch — src/models/recommendation_engine.py imports the missing src.models.integrated_selector (#21), so that module is unimportable until #21 is resolved. That edit is a defensive guard for when #21 lands; nothing here depends on #1/PR #19.