feat(infra): add sweep coordinator, structured logging, test suite, and README

Sweep coordinator (Step 16):
- SweepCoordinator runs all 8 agents in parallel with 60s per-agent / 300s total timeout
- Aggregates findings, actions, errors into SweepCoordinatorResult
- Registered in FastAPI lifespan; triggered via POST /sweep

Structured logging (Step 18):
- logging_utils/structured.py: JSONFormatter emitting ts/level/logger/msg + custom fields
- log_directive_event() for structured directive lifecycle logging
- push_to_loki() async Loki push (graceful no-op if LOKI_URL unset)
- configure_logging() replaces root handler at startup

Tests (Steps 17+19):
- conftest.py: mock_odoo, mock_pool, mock_llm fixtures
- test_tool_validator.py: 9 tests covering validation, coercion, hallucination stripping
- test_llm_router.py: 6 tests covering local/cloud/hybrid modes and HIPAA enforcement
- test_peer_bus.py: 6 tests covering registration, timeout, depth, circular detection
- test_finance_agent.py: 10 tests covering all 6 steps + sweep + peer request
- test_memory_manager.py: 3 tests covering context build + hard cap enforcement
- test_dispatch_router.py: 3 tests covering dispatch, rate limit, health endpoint
- test_odoo_client.py: 4 tests covering search_read, write result, unlink warning
- test_e2e_dispatch.py: 2 E2E tests - full dispatch cycle + peer bus communication

README (Step 20): architecture diagram, privacy modes, quick start, env vars, structure

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

tests/test_tool_validator.py

@@ -0,0 +1,73 @@
+import pytest
+from agent_service.llm.tool_validator import ToolCallValidator, AgentConfigError
+
+SAMPLE_TOOLS = [
+    {'name': 'get_invoices', 'parameters': {
+        'state': {'type': 'string', 'optional': True},
+        'limit': {'type': 'integer', 'optional': True},
+        'partner_id': {'type': 'integer'},
+    }},
+    {'name': 'send_reminder', 'parameters': {
+        'invoice_id': {'type': 'integer'},
+        'message': {'type': 'string', 'optional': True},
+    }},
+]
+
+
+def test_validator_init():
+    v = ToolCallValidator(SAMPLE_TOOLS)
+    assert 'get_invoices' in v._tool_map
+
+
+def test_raises_on_too_many_tools():
+    many_tools = [{'name': f'tool_{i}', 'parameters': {}} for i in range(9)]
+    with pytest.raises(AgentConfigError, match='MAX_TOOLS_PER_AGENT'):
+        ToolCallValidator(many_tools)
+
+
+def test_valid_tool_call():
+    v = ToolCallValidator(SAMPLE_TOOLS)
+    result = v.validate({'name': 'get_invoices', 'arguments': {'partner_id': 42}})
+    assert result.name == 'get_invoices'
+    assert result.arguments['partner_id'] == 42
+
+
+def test_strips_hallucinated_params():
+    v = ToolCallValidator(SAMPLE_TOOLS)
+    result = v.validate({'name': 'get_invoices', 'arguments': {
+        'partner_id': 1, 'nonexistent_param': 'bad',
+    }})
+    assert 'nonexistent_param' not in result.arguments
+
+
+def test_missing_required_param_raises():
+    v = ToolCallValidator(SAMPLE_TOOLS)
+    with pytest.raises(ValueError, match='partner_id'):
+        v.validate({'name': 'get_invoices', 'arguments': {}})
+
+
+def test_type_coercion_int():
+    v = ToolCallValidator(SAMPLE_TOOLS)
+    result = v.validate({'name': 'get_invoices', 'arguments': {'partner_id': '42'}})
+    assert result.arguments['partner_id'] == 42
+
+
+def test_unknown_tool_raises():
+    v = ToolCallValidator(SAMPLE_TOOLS)
+    with pytest.raises(ValueError, match='Unknown tool'):
+        v.validate({'name': 'nonexistent_tool', 'arguments': {}})
+
+
+def test_parse_or_fallback_returns_none_on_bad_json():
+    v = ToolCallValidator(SAMPLE_TOOLS)
+    result = v.parse_or_fallback('not json at all', context='test')
+    assert result is None
+
+
+def test_parse_or_fallback_valid_json():
+    v = ToolCallValidator(SAMPLE_TOOLS)
+    import json
+    raw = json.dumps({'name': 'send_reminder', 'arguments': {'invoice_id': 5}})
+    result = v.parse_or_fallback(raw, context='test')
+    assert result is not None
+    assert result.name == 'send_reminder'