feat(07-03): AnthropicProvider singleton + output_config + truncation — D-03/D-07/D-12/D-13

- Delete MAX_AI_CHARS constant and _client() method from anthropic_provider.py
- Add self._client = AsyncAnthropic(...) singleton in __init__ (D-07)
- Add _truncate() with 60/40 split using self._context_chars (D-13)
- Add _CLASSIFICATION_SCHEMA and _SUGGESTIONS_SCHEMA module-level constants
- classify() and suggest_topics() pass output_config with json_schema format (D-03)
- stop_reason != "end_turn" degrades to parse_classification("") (T-07-08)
- Widen __init__ signature to (api_key, model, context_chars, base_url) (uniform ctor)
- Update ai/__init__.py to pass context_chars + base_url to AnthropicProvider
- Promote test_anthropic_structured_output + add test_anthropic_stop_reason_fallback

backend/tests/test_ai_providers.py

@@ -5,11 +5,12 @@ Wave 2 (Plan 07-02) promotes: test_get_provider_typed, test_client_singleton,
 test_generic_openai_json_mode, test_context_chars_truncation, test_smart_truncation,
 test_gemini_fallback_to_parse_classification.
 
-Remaining stubs (promoted in Plan 07-03): test_anthropic_structured_output.
+Wave 3 (Plan 07-03) promotes: test_anthropic_structured_output.
 """
 import pytest
 from unittest.mock import AsyncMock, MagicMock, patch
 
+from ai.anthropic_provider import AnthropicProvider, _CLASSIFICATION_SCHEMA
 from ai.generic_openai_provider import GenericOpenAIProvider
 from ai.openai_provider import OpenAIProvider
 from ai.provider_config import ProviderConfig, PROVIDER_DEFAULTS
@@ -189,9 +190,81 @@ async def test_gemini_fallback_to_parse_classification():
 
 
 # ---------------------------------------------------------------------------
-# Stub: promoted in Plan 07-03
+# Task 1 (Plan 07-03): AnthropicProvider output_config structured output — D-03
 # ---------------------------------------------------------------------------
 
-@pytest.mark.xfail(strict=False, reason="Wave 0 stub — promoted in Plan 07-03")
+@pytest.mark.asyncio
 async def test_anthropic_structured_output():
-    pytest.xfail("not implemented yet — Plan 07-03")
+    """AnthropicProvider.classify() passes output_config with the classification schema (D-03).
+
+    Verifies:
+    - output_config kwarg is present in the messages.create call
+    - output_config value matches _CLASSIFICATION_SCHEMA exactly
+    - Provider correctly reads response.content[0].text when stop_reason == "end_turn"
+    - Singleton _client is reused (AsyncAnthropic constructed once per provider instance)
+    """
+    provider = AnthropicProvider(api_key="test-key", model="claude-sonnet-4-6", context_chars=100)
+
+    # Build a stub response that models a successful end_turn response
+    stub_content = MagicMock()
+    stub_content.text = '{"assigned_topics":[],"new_topic_suggestions":[]}'
+    stub_response = MagicMock()
+    stub_response.content = [stub_content]
+    stub_response.stop_reason = "end_turn"
+
+    mock_create = AsyncMock(return_value=stub_response)
+
+    with patch("ai.anthropic_provider.anthropic.AsyncAnthropic") as mock_cls:
+        mock_client = MagicMock()
+        mock_client.messages = MagicMock()
+        mock_client.messages.create = mock_create
+        mock_cls.return_value = mock_client
+
+        # Re-create provider inside the patch so self._client uses the mock
+        provider = AnthropicProvider(
+            api_key="test-key", model="claude-sonnet-4-6", context_chars=100
+        )
+        result = await provider.classify("short doc text", [], "sys prompt")
+
+    # output_config must be present and match the schema (D-03)
+    call_kwargs = mock_create.await_args.kwargs
+    assert "output_config" in call_kwargs, "output_config must be passed to messages.create()"
+    assert call_kwargs["output_config"] == {
+        "format": {"type": "json_schema", "schema": _CLASSIFICATION_SCHEMA}
+    }, "output_config must use _CLASSIFICATION_SCHEMA"
+
+    # AsyncAnthropic must have been constructed exactly once (D-07 singleton)
+    assert mock_cls.call_count == 1, "AsyncAnthropic must be constructed once (singleton)"
+
+    # Result must be a valid ClassificationResult
+    assert result.topics == []
+    assert result.suggested_new_topics == []
+
+
+@pytest.mark.asyncio
+async def test_anthropic_stop_reason_fallback():
+    """When stop_reason is not 'end_turn', AnthropicProvider falls back to empty ClassificationResult.
+
+    T-07-08: refusal or max_tokens stop_reason must not crash — parse_classification("") returns
+    an empty result rather than raising an exception.
+    """
+    stub_content = MagicMock()
+    stub_content.text = '{"assigned_topics":["should","be","ignored"]}'
+    stub_response = MagicMock()
+    stub_response.content = [stub_content]
+    stub_response.stop_reason = "max_tokens"  # simulated refusal / truncation
+
+    mock_create = AsyncMock(return_value=stub_response)
+
+    with patch("ai.anthropic_provider.anthropic.AsyncAnthropic") as mock_cls:
+        mock_client = MagicMock()
+        mock_client.messages = MagicMock()
+        mock_client.messages.create = mock_create
+        mock_cls.return_value = mock_client
+
+        provider = AnthropicProvider(api_key="k", model="claude-sonnet-4-6", context_chars=1000)
+        result = await provider.classify("doc text", [], "sys")
+
+    # stop_reason != "end_turn" → raw == "" → empty ClassificationResult, no crash
+    assert result.topics == []
+    assert result.suggested_new_topics == []