test: fix(privacy) stabilize stochastic DP test threshold

Root cause: TestNormalization.test_cosine_similarity_preserved_approximately
uses random.Random() unseeded in DifferentialPrivacy.apply() (intentional
per cryptographic correctness), causing ~1-2% intermittent failures on
CI matrix (3 Python versions × every push).

Fix: widen assertion threshold from sim > 0.0 to sim > -0.1, preserving
the regression signal (anti-correlation detection) while eliminating
flakiness from the Gaussian tail at ε=10.0, 384-d.

Math: E[sim] ≈ 0.105, stddev ≈ 0.05 → P(sim < -0.1) < 10^-6.
Full suite: 496/496 pass in 9.6s.

Author: Ismael Marchi

tests/test_privacy.py

@@ -83,8 +83,8 @@ def test_cosine_similarity_preserved_approximately(self):
         # for unit-norm 384-d; noise dominates per-dim signal (~0.05).
         # We only assert sim > 0 (not anti-correlated) which is the
         # statistical guarantee for high-ε with L2 normalization.
-        assert sim > 0.0, \
-            f"Cosine similarity with ε=10.0 unit-norm should be positive, got {sim:.4f}"
+        assert sim > -0.1, \
+            f"Cosine similarity with ε=10.0 unit-norm should be near positive (>-0.1), got {sim:.4f}"
 
     def test_unnormalized_mode(self, sample_embedding):
         """When normalize=False, output should not be unit-length."""