feat(logic): dostawcy LLM i pisanie horoskopu (LOG-31)

Domyslnie model LOKALNY — prompt niesie oryginalne opisy z baz, wiec
domyslnie NIC nie opuszcza sieci. Chmura wlaczana swiadomie (LOG-32).

- app/llm/: LLMProvider (jak EphemerisEngine z LOG-24) + dwie implementacje.
  Lokalny serwer modelu (Ollama/vLLM/llama.cpp) i OpenAI mowia TYM SAMYM
  protokolem /chat/completions, wiec obsluguje je jedna klasa; Anthropic ma
  wlasny /v1/messages. Napisane na samym httpx — bez SDK openai/anthropic:
  mniej zaleznosci i pelna kontrola nad tym, co wychodzi z sieci.
- Kazdy dostawca deklaruje `leaves_lan` — interfejs MUSI jawnie mowic, czy
  tresc baz opuszcza siec; UI na tej podstawie ostrzega.
- Ponawianie z backoffem (429/5xx), timeouty, czytelne bledy zamiast stacktrace.
- Klucz wylacznie z LLM_API_KEY (sekret), nigdy w repo ani w UI.
- POST /chart/horoscope + GET /llm/health.

WAZNE: prompt jest zwracany ZAWSZE — takze gdy model padnie lub brakuje
klucza. Dzieki temu awaria dostawcy nie blokuje pracy: prompt mozna
skopiowac i uzyc recznie.

Prezentacja: wybor modelu (lokalny/Anthropic/OpenAI), przycisk „Napisz
horoskop (AI)", wynik z informacja kto go napisal, czy dane opuscily siec,
ile wskazan weszlo, oraz zastrzezenie ze to nie porada medyczna (PRE-15).

Testy: 13 nowych (transport podstawiony — zaden prawdziwy model nie wolany),
calosc 131 passed / 1 skipped. Zweryfikowane e2e na atrapie serwera modelu:
horoskop napisany, leaves_lan=false, tokeny zliczone; a przy padnietym
modelu / braku klucza / zlym dostawcy — czytelny blad i zachowany prompt.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

services/logic/tests/test_llm.py

@@ -0,0 +1,149 @@
+"""Dostawcy LLM (LOG-31) — bez wołania jakiegokolwiek prawdziwego modelu.
+
+Transport podstawiamy przez httpx.MockTransport, więc testy są szybkie,
+deterministyczne i nic nie wychodzi na zewnątrz.
+"""
+import httpx
+import pytest
+
+from app.llm import factory
+from app.llm.base import Completion, LLMError
+from app.llm.providers import AnthropicProvider, ChatCompletionsProvider
+
+
+def _mock_client(handler):
+    """Podmienia httpx.Client na wersję z transportem testowym."""
+    class _C(httpx.Client):
+        def __init__(self, *a, **kw):
+            kw["transport"] = httpx.MockTransport(handler)
+            super().__init__(*a, **kw)
+    return _C
+
+
+@pytest.fixture
+def chat_ok(monkeypatch):
+    def handler(request):
+        assert request.url.path.endswith("/chat/completions")
+        return httpx.Response(200, json={
+            "model": "test-model",
+            "choices": [{"message": {"content": "Horoskop testowy."}}],
+            "usage": {"prompt_tokens": 100, "completion_tokens": 50},
+        })
+    monkeypatch.setattr(httpx, "Client", _mock_client(handler))
+
+
+# ------------------------------------------------------- protokół chat/completions
+
+def test_local_provider_generates(chat_ok):
+    p = ChatCompletionsProvider("local", "http://localhost:11434/v1", "m", leaves_lan=False)
+    out = p.generate("prompt", 500)
+    assert isinstance(out, Completion)
+    assert out.text == "Horoskop testowy."
+    assert out.usage["completion_tokens"] == 50
+
+
+def test_local_provider_does_not_leave_lan(chat_ok):
+    p = ChatCompletionsProvider("local", "http://localhost:11434/v1", "m", leaves_lan=False)
+    assert p.generate("prompt", 100).leaves_lan is False
+
+
+def test_cloud_provider_marks_leaving_lan(chat_ok):
+    p = ChatCompletionsProvider("openai", "https://api.openai.com/v1", "m", "klucz",
+                                leaves_lan=True)
+    assert p.generate("prompt", 100).leaves_lan is True
+
+
+def test_api_key_sent_only_when_set(monkeypatch):
+    seen = {}
+
+    def handler(request):
+        seen["auth"] = request.headers.get("authorization")
+        return httpx.Response(200, json={"choices": [{"message": {"content": "x"}}]})
+
+    monkeypatch.setattr(httpx, "Client", _mock_client(handler))
+    ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 10)
+    assert seen["auth"] is None
+    ChatCompletionsProvider("openai", "http://x/v1", "m", "tajny").generate("p", 10)
+    assert seen["auth"] == "Bearer tajny"
+
+
+def test_http_error_becomes_readable_message(monkeypatch):
+    monkeypatch.setattr(httpx, "Client",
+                        _mock_client(lambda r: httpx.Response(400, text="zly model")))
+    with pytest.raises(LLMError, match="400"):
+        ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 10)
+
+
+def test_malformed_response_reported(monkeypatch):
+    monkeypatch.setattr(httpx, "Client",
+                        _mock_client(lambda r: httpx.Response(200, json={"nonsens": 1})))
+    with pytest.raises(LLMError, match="kształt"):
+        ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 10)
+
+
+def test_retries_then_succeeds(monkeypatch):
+    calls = {"n": 0}
+
+    def handler(request):
+        calls["n"] += 1
+        if calls["n"] < 3:
+            return httpx.Response(429, text="za duzo")
+        return httpx.Response(200, json={"choices": [{"message": {"content": "ok"}}]})
+
+    monkeypatch.setattr(httpx, "Client", _mock_client(handler))
+    monkeypatch.setattr("app.llm.providers.time.sleep", lambda s: None)   # bez czekania
+    assert ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 10).text == "ok"
+    assert calls["n"] == 3
+
+
+# ------------------------------------------------------------------- Anthropic
+
+def test_anthropic_generates(monkeypatch):
+    def handler(request):
+        assert request.url.path.endswith("/v1/messages")
+        assert request.headers.get("x-api-key") == "klucz"
+        assert request.headers.get("anthropic-version")
+        return httpx.Response(200, json={
+            "model": "claude-x",
+            "content": [{"type": "text", "text": "Prognoza."}],
+            "usage": {"input_tokens": 10, "output_tokens": 5},
+        })
+
+    monkeypatch.setattr(httpx, "Client", _mock_client(handler))
+    out = AnthropicProvider("https://api.anthropic.com", "claude-x", "klucz").generate("p", 100)
+    assert out.text == "Prognoza." and out.leaves_lan is True
+
+
+def test_anthropic_requires_key():
+    with pytest.raises(LLMError, match="LLM_API_KEY"):
+        AnthropicProvider("https://api.anthropic.com", "m", "").generate("p", 10)
+
+
+# --------------------------------------------------------------------- fabryka
+
+def test_default_provider_is_local(monkeypatch):
+    monkeypatch.delenv("LLM_PROVIDER", raising=False)
+    monkeypatch.delenv("LLM_MODEL", raising=False)
+    monkeypatch.delenv("LLM_BASE_URL", raising=False)
+    p = factory.build_provider()
+    assert p.name == "local"
+    # domyślnie NIC nie opuszcza sieci — prompt niesie opisy z baz (LOG-32)
+    assert p.leaves_lan is False
+
+
+def test_openai_requires_key(monkeypatch):
+    monkeypatch.delenv("LLM_API_KEY", raising=False)
+    with pytest.raises(LLMError, match="LLM_API_KEY"):
+        factory.build_provider("openai")
+
+
+def test_unknown_provider_rejected():
+    with pytest.raises(LLMError, match="Nieznany dostawca"):
+        factory.build_provider("bzdura")
+
+
+def test_env_overrides_model_and_url(monkeypatch):
+    monkeypatch.setenv("LLM_MODEL", "moj-model")
+    monkeypatch.setenv("LLM_BASE_URL", "http://serwer:8000/v1")
+    p = factory.build_provider("local")
+    assert p.model == "moj-model" and p.base_url == "http://serwer:8000/v1"