feat(mcp-docugen): Task 4-6 template_store, llm_client, generation_store

- TemplateStore: CRUD filesystem + asset dir, frontmatter YAML roundtrip,
  path traversal rejection
- OpenRouterClient: async httpx con retry backoff esponenziale (5xx, 429,
  timeout), no-retry su 4xx, parse usage/cost
- GenerationStore: SQLite aiosqlite con schema generations + ephemeral_assets,
  cleanup TTL, stats aggregate

Root pyproject aggiornato con respx + pytest-cov dev deps.
19 + 11 + 9 + 6 = 45 test totali, tutti passed.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

services/mcp-docugen/tests/unit/test_llm_client.py

@@ -0,0 +1,173 @@
+import httpx
+import pytest
+import respx
+
+from mcp_docugen.llm_client import (
+    LLMAuthError,
+    LLMEmptyResponse,
+    LLMInvalidResponse,
+    LLMRateLimit,
+    LLMTimeout,
+    LLMUpstreamError,
+    OpenRouterClient,
+)
+
+
+def _success_body(text: str = "output text") -> dict:
+    return {
+        "id": "gen-1",
+        "choices": [{"message": {"role": "assistant", "content": text}}],
+        "model": "anthropic/claude-sonnet-4",
+        "usage": {
+            "prompt_tokens": 100,
+            "completion_tokens": 200,
+            "total_cost": 0.01,
+        },
+    }
+
+
+@respx.mock
+async def test_chat_success():
+    respx.post("https://openrouter.ai/api/v1/chat/completions").mock(
+        return_value=httpx.Response(200, json=_success_body("hello"))
+    )
+    client = OpenRouterClient(
+        api_key="sk", base_url="https://openrouter.ai/api/v1", timeout=5
+    )
+    resp = await client.chat(
+        model="anthropic/claude-sonnet-4", system="sys", user="user"
+    )
+    assert resp.text == "hello"
+    assert resp.tokens_in == 100
+    assert resp.tokens_out == 200
+    assert resp.cost_usd == 0.01
+    assert resp.model == "anthropic/claude-sonnet-4"
+
+
+@respx.mock
+async def test_chat_retries_on_5xx():
+    route = respx.post("https://openrouter.ai/api/v1/chat/completions").mock(
+        side_effect=[
+            httpx.Response(503),
+            httpx.Response(502),
+            httpx.Response(200, json=_success_body()),
+        ]
+    )
+    client = OpenRouterClient(
+        api_key="sk",
+        base_url="https://openrouter.ai/api/v1",
+        timeout=5,
+        retry_base_delay=0,
+    )
+    resp = await client.chat(model="m", system="s", user="u")
+    assert resp.text == "output text"
+    assert route.call_count == 3
+
+
+@respx.mock
+async def test_chat_exhausts_retries_5xx():
+    respx.post("https://openrouter.ai/api/v1/chat/completions").mock(
+        return_value=httpx.Response(500)
+    )
+    client = OpenRouterClient(
+        api_key="sk",
+        base_url="https://openrouter.ai/api/v1",
+        timeout=5,
+        retry_base_delay=0,
+    )
+    with pytest.raises(LLMUpstreamError):
+        await client.chat(model="m", system="s", user="u")
+
+
+@respx.mock
+async def test_chat_retries_on_429():
+    route = respx.post("https://openrouter.ai/api/v1/chat/completions").mock(
+        side_effect=[
+            httpx.Response(429),
+            httpx.Response(200, json=_success_body()),
+        ]
+    )
+    client = OpenRouterClient(
+        api_key="sk",
+        base_url="https://openrouter.ai/api/v1",
+        timeout=5,
+        retry_base_delay=0,
+    )
+    resp = await client.chat(model="m", system="s", user="u")
+    assert route.call_count == 2
+    assert resp.text == "output text"
+
+
+@respx.mock
+async def test_chat_exhausts_retries_429():
+    respx.post("https://openrouter.ai/api/v1/chat/completions").mock(
+        return_value=httpx.Response(429)
+    )
+    client = OpenRouterClient(
+        api_key="sk",
+        base_url="https://openrouter.ai/api/v1",
+        timeout=5,
+        retry_base_delay=0,
+    )
+    with pytest.raises(LLMRateLimit):
+        await client.chat(model="m", system="s", user="u")
+
+
+@respx.mock
+async def test_chat_no_retry_on_401():
+    respx.post("https://openrouter.ai/api/v1/chat/completions").mock(
+        return_value=httpx.Response(401)
+    )
+    client = OpenRouterClient(
+        api_key="sk",
+        base_url="https://openrouter.ai/api/v1",
+        timeout=5,
+        retry_base_delay=0,
+    )
+    with pytest.raises(LLMAuthError):
+        await client.chat(model="m", system="s", user="u")
+
+
+@respx.mock
+async def test_chat_timeout():
+    respx.post("https://openrouter.ai/api/v1/chat/completions").mock(
+        side_effect=httpx.ReadTimeout("timeout")
+    )
+    client = OpenRouterClient(
+        api_key="sk",
+        base_url="https://openrouter.ai/api/v1",
+        timeout=1,
+        retry_base_delay=0,
+    )
+    with pytest.raises(LLMTimeout):
+        await client.chat(model="m", system="s", user="u")
+
+
+@respx.mock
+async def test_chat_invalid_response_shape():
+    respx.post("https://openrouter.ai/api/v1/chat/completions").mock(
+        return_value=httpx.Response(200, json={"no": "choices"})
+    )
+    client = OpenRouterClient(
+        api_key="sk",
+        base_url="https://openrouter.ai/api/v1",
+        timeout=5,
+        retry_base_delay=0,
+    )
+    with pytest.raises(LLMInvalidResponse):
+        await client.chat(model="m", system="s", user="u")
+
+
+@respx.mock
+async def test_chat_empty_content():
+    respx.post("https://openrouter.ai/api/v1/chat/completions").mock(
+        return_value=httpx.Response(200, json=_success_body(text=""))
+    )
+    client = OpenRouterClient(
+        api_key="sk",
+        base_url="https://openrouter.ai/api/v1",
+        timeout=5,
+        retry_base_delay=0,
+    )
+    with pytest.raises(LLMEmptyResponse):
+        await client.chat(model="m", system="s", user="u")