fix probability over-confidence: model target-before-stop, not just touch

Backtest (32k setups) showed the touch-only probability model was ~2x
over-confident — predicted 70% hit 39%, predicted 88% hit 46% — because it
ignored the competing stop. estimate_probability now multiplies the reach
probability (touch within horizon) by the two-barrier gambler's-ruin ratio
1/(R:R+1) = P(target before stop). A 3:1 setup now reads ~25% base, not ~70%,
which lines up with realized rates. Strength/alignment modulation unchanged.

Recalibrates every probability and the EV ranking; the min_target_probability
gate threshold now means roughly what it says. Re-run the backtest to confirm
the calibration table flattens toward the diagonal.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

tests/unit/test_recommendation_service.py

@@ -205,9 +205,10 @@ def test_probability_decreases_with_distance():
 
     # Monotonic decay with distance
     assert near > mid > far
-    # Near target is genuinely likely; a 10-ATR target is a long shot
-    assert near > 60
-    assert far < 25
+    # Backtest-calibrated: even a near target with no R:R context (even race) is
+    # only a moderate probability, and a 10-ATR target is a long shot.
+    assert near > 30
+    assert far < 15
 
 
 def test_far_target_not_high_probability_even_with_strong_level():