feat(training): add Timmy voice batch 09 dataset (#589)

Generate a deterministic batch09 Timmy voice corpus with 1,000 ShareGPT prompt-response pairs,
50 approved source sessions, a source manifest, and focused validation tests.

tests/test_timmy_voice_batch09.py

@@ -0,0 +1,126 @@
+import json
+import re
+from collections import Counter, defaultdict
+from pathlib import Path
+
+DATA_FILE = Path('training-data/timmy-voice-batch09.jsonl')
+SCRIPT_FILE = Path('training-data/generate_timmy_voice_batch09.py')
+SOURCES_FILE = Path('training-data/timmy-voice-batch09.sources.json')
+README_FILE = Path('training-data/README-batch09.md')
+EXPECTED_ROWS = 1000
+EXPECTED_SOURCE_SESSIONS = 50
+EXPECTED_ROWS_PER_SESSION = 20
+BANNED_MODEL_STRINGS = ['claude', 'gpt-4', 'gpt-3', 'gpt', 'o1', 'o3', 'gemini', 'anthropic', 'openai', 'unknown']
+CRISIS_TERMS = ['suicide', 'kill myself', 'end my life', 'overdose', 'bridge', 'gun', 'die']
+CREDENTIAL_MARKERS = ['password', 'pass:', 'token', 'api key', 'secret', 'login:']
+PRIVATE_MARKERS = ['/users/', '/private/', '~/.timmy', '~/.hermes', 'alexanderwhitestone.com']
+META_MARKERS = [
+    'review the conversation above',
+    "you've reached the maximum number of tool-calling iterations",
+    'without calling any more tools',
+    'nothing to save',
+    '[silent]',
+]
+EMAIL_RE = re.compile(r'[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}')
+
+
+def load_rows(path: Path):
+    assert path.exists(), f'missing file: {path}'
+    rows = []
+    for line in path.read_text(encoding='utf-8').splitlines():
+        line = line.strip()
+        if line:
+            rows.append(json.loads(line))
+    return rows
+
+
+def load_sources():
+    assert SOURCES_FILE.exists(), f'missing file: {SOURCES_FILE}'
+    return json.loads(SOURCES_FILE.read_text(encoding='utf-8'))
+
+
+def test_batch09_artifacts_exist():
+    for path in [DATA_FILE, SCRIPT_FILE, SOURCES_FILE, README_FILE]:
+        assert path.exists(), f'missing artifact: {path}'
+
+
+def test_sources_manifest_has_50_unique_sessions():
+    manifest = load_sources()
+    assert manifest['batch'] == 9
+    assert manifest['total_source_sessions'] == EXPECTED_SOURCE_SESSIONS
+    sessions = manifest['sessions']
+    assert len(sessions) == EXPECTED_SOURCE_SESSIONS
+    session_files = [item['session_file'] for item in sessions]
+    assert len(session_files) == len(set(session_files))
+    for item in sessions:
+        assert item['session_score'] > 0
+        assert item['pair_score'] > 0
+        assert item['source_prompt'].strip()
+        assert item['source_response'].strip()
+        model = item['model'].lower()
+        assert not any(bad in model for bad in BANNED_MODEL_STRINGS), item['model']
+
+
+def test_dataset_has_exactly_1000_rows():
+    rows = load_rows(DATA_FILE)
+    assert len(rows) == EXPECTED_ROWS
+
+
+def test_ids_are_unique_and_batch_scoped():
+    rows = load_rows(DATA_FILE)
+    ids = [row['id'] for row in rows]
+    assert len(ids) == len(set(ids))
+    assert ids[0] == 'timmy-voice-batch09-0001'
+    assert ids[-1] == 'timmy-voice-batch09-1000'
+    for row in rows:
+        assert row['batch'] == 9
+        assert row['model'] == 'timmy-voice-batch09'
+        assert row['quality_score'] >= 0.80
+
+
+def test_sharegpt_schema_and_source_mapping():
+    rows = load_rows(DATA_FILE)
+    manifest = load_sources()
+    valid_sources = {item['session_file'] for item in manifest['sessions']}
+    counts = Counter()
+    prompts_by_source = defaultdict(set)
+    for row in rows:
+        assert row['source'] == 'session_derived_approved'
+        assert row['source_session'] in valid_sources
+        counts[row['source_session']] += 1
+        conv = row['conversations']
+        assert [item['from'] for item in conv] == ['system', 'human', 'gpt']
+        assert 'Sovereignty and service always.' in conv[0]['value']
+        assert conv[1]['value'].strip()
+        assert conv[2]['value'].strip()
+        prompts_by_source[row['source_session']].add(conv[1]['value'])
+    assert len(counts) == EXPECTED_SOURCE_SESSIONS
+    assert set(counts.values()) == {EXPECTED_ROWS_PER_SESSION}
+    for source_session, prompts in prompts_by_source.items():
+        assert len(prompts) == EXPECTED_ROWS_PER_SESSION, source_session
+
+
+def test_crisis_prompts_include_required_protocol():
+    rows = load_rows(DATA_FILE)
+    for row in rows:
+        prompt = row['conversations'][1]['value'].lower()
+        response = row['conversations'][2]['value']
+        if any(term in prompt for term in CRISIS_TERMS):
+            assert 'Are you safe right now?' in response
+            assert '988' in response
+
+
+def test_no_credentials_or_private_paths_leak_into_sources_or_rows():
+    manifest = load_sources()
+    corpus = json.dumps(manifest).lower() + '\n' + DATA_FILE.read_text(encoding='utf-8').lower()
+    assert not any(marker in corpus for marker in CREDENTIAL_MARKERS)
+    assert not any(marker in corpus for marker in PRIVATE_MARKERS)
+    assert not any(marker in corpus for marker in META_MARKERS)
+    assert not EMAIL_RE.search(corpus)
+
+
+def test_readme_documents_generation_contract():
+    text = README_FILE.read_text(encoding='utf-8')
+    assert '50 source sessions' in text
+    assert '20 prompt variations per session' in text
+    assert 'approved-model provenance filter' in text