feat: voice recognition — resemblyzer embed + cosine similarity matching

Implements pack/unpack_embedding, cosine_similarity, find_best_speaker_match
(max-per-speaker grouping, known/ambiguous/unknown thresholds), and embed_audio
with lazy resemblyzer import so tests run without the docker-only dependency.
10 tests passing.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

tests/test_recognition.py

@@ -0,0 +1,106 @@
+import numpy as np
+import struct
+import pytest
+from unittest.mock import patch, MagicMock
+
+
+def test_pack_unpack_roundtrip():
+    from app.pipeline.recognition import pack_embedding, unpack_embedding
+    original = np.array([0.1, 0.2, 0.3, -0.5], dtype=np.float32)
+    packed = pack_embedding(original)
+    recovered = unpack_embedding(packed)
+    np.testing.assert_allclose(recovered, original, rtol=1e-6)
+
+
+def test_cosine_similarity_identical():
+    from app.pipeline.recognition import cosine_similarity
+    v = np.array([1.0, 0.0, 0.0], dtype=np.float32)
+    assert cosine_similarity(v, v) == pytest.approx(1.0)
+
+
+def test_cosine_similarity_orthogonal():
+    from app.pipeline.recognition import cosine_similarity
+    a = np.array([1.0, 0.0], dtype=np.float32)
+    b = np.array([0.0, 1.0], dtype=np.float32)
+    assert cosine_similarity(a, b) == pytest.approx(0.0)
+
+
+def test_cosine_similarity_zero_vector():
+    from app.pipeline.recognition import cosine_similarity
+    a = np.array([0.0, 0.0], dtype=np.float32)
+    b = np.array([1.0, 0.0], dtype=np.float32)
+    assert cosine_similarity(a, b) == 0.0
+
+
+def test_find_best_match_known():
+    from app.pipeline.recognition import find_best_speaker_match, pack_embedding
+    ref = np.array([1.0, 0.0, 0.0], dtype=np.float32)
+    query = np.array([0.98, 0.2, 0.0], dtype=np.float32)
+    query /= np.linalg.norm(query)
+    ref /= np.linalg.norm(ref)
+    status, sid, conf = find_best_speaker_match(
+        query, [(1, pack_embedding(ref))], known_threshold=0.85, ambiguous_threshold=0.60
+    )
+    assert status == "known"
+    assert sid == 1
+    assert conf >= 0.85
+
+
+def test_find_best_match_ambiguous():
+    from app.pipeline.recognition import find_best_speaker_match, pack_embedding
+    ref = np.array([1.0, 0.0, 0.0], dtype=np.float32)
+    # cos similarity ~0.707 (45 degrees)
+    query = np.array([1.0, 1.0, 0.0], dtype=np.float32)
+    query /= np.linalg.norm(query)
+    ref /= np.linalg.norm(ref)
+    status, sid, conf = find_best_speaker_match(
+        query, [(1, pack_embedding(ref))], known_threshold=0.85, ambiguous_threshold=0.60
+    )
+    assert status == "ambiguous"
+    assert sid == 1
+
+
+def test_find_best_match_unknown():
+    from app.pipeline.recognition import find_best_speaker_match, pack_embedding
+    ref = np.array([1.0, 0.0, 0.0], dtype=np.float32)
+    query = np.array([0.0, 1.0, 0.0], dtype=np.float32)  # orthogonal = 0 similarity
+    status, sid, conf = find_best_speaker_match(
+        query, [(1, pack_embedding(ref))], known_threshold=0.85, ambiguous_threshold=0.60
+    )
+    assert status == "unknown"
+    assert sid is None
+
+
+def test_find_best_match_empty_embeddings():
+    from app.pipeline.recognition import find_best_speaker_match
+    query = np.array([1.0, 0.0], dtype=np.float32)
+    status, sid, conf = find_best_speaker_match(query, [])
+    assert status == "unknown"
+    assert sid is None
+
+
+def test_find_best_match_takes_max_per_speaker():
+    from app.pipeline.recognition import find_best_speaker_match, pack_embedding
+    # Speaker 1 has two embeddings — one poor, one good
+    ref_bad = np.array([0.0, 1.0, 0.0], dtype=np.float32)
+    ref_good = np.array([1.0, 0.0, 0.0], dtype=np.float32)
+    query = np.array([1.0, 0.0, 0.0], dtype=np.float32)
+    stored = [(1, pack_embedding(ref_bad)), (1, pack_embedding(ref_good))]
+    status, sid, conf = find_best_speaker_match(
+        query, stored, known_threshold=0.85, ambiguous_threshold=0.60
+    )
+    assert status == "known"
+    assert sid == 1
+
+
+def test_embed_audio_calls_resemblyzer(mocker):
+    from app.pipeline.recognition import embed_audio
+    mock_wav = np.zeros(16000, dtype=np.float32)
+    mock_emb = np.ones(256, dtype=np.float32)
+    mocker.patch("app.pipeline.recognition.preprocess_wav", return_value=mock_wav)
+    mock_encoder = MagicMock()
+    mock_encoder.embed_utterance.return_value = mock_emb
+    mocker.patch("app.pipeline.recognition.get_encoder", return_value=mock_encoder)
+    result = embed_audio(b"\x00" * 100)
+    assert result.shape == (256,)
+    mock_encoder.embed_utterance.assert_called_once_with(mock_wav)