Refacto du code coté Python, java et coté angular afin de mieux séparer les responsabilité et d'avoir moins de répétitivité dans le code.

Mise en place de tests unitaires coté Python et Angular
Mise en place de la couverture de test directement dans le workflow : le programme ne build pas si jamais un test échoue
Passage en v0.16.2 en conséquence

brain/.coveragerc

@@ -0,0 +1,15 @@
+# Configuration de couverture (coverage.py / pytest-cov).
+# Rapport HTML (équivalent JaCoCo) : pytest --cov=app --cov-report=html → htmlcov/
+# Plancher anti-régression appliqué en CI : --cov-fail-under=50
+[run]
+source = app
+branch = false
+
+[report]
+show_missing = true
+skip_covered = false
+# Lignes jamais comptées comme « à couvrir ».
+exclude_lines =
+    pragma: no cover
+    if __name__ == .__main__.:
+    raise NotImplementedError

brain/.gitignore

@@ -2,3 +2,8 @@
 __pycache__/
 *.pyc
 .env
+
+# Couverture de tests (pytest-cov)
+htmlcov/
+.coverage
+.pytest_cache/

brain/app/infrastructure/base_openai_adapter.py

@@ -0,0 +1,226 @@
+"""Socle commun aux adapters LLM « OpenAI-compatible » (OpenRouter, Gemini,
+Mistral) — ils exposent tous `POST {base}/chat/completions` en SSE avec le même
+schéma de payload et de flux.
+
+Cette classe de base porte la mécanique partagée (construction du payload, appel
+HTTP streamé, parsing SSE, garde-fous de timeout au temps écoulé, traduction des
+erreurs). Chaque adapter concret ne fournit plus que ses spécificités :
+URL, en-têtes, support du mode JSON natif, messages d'erreur, lecture de la config.
+
+`generate` one-shot passe lui aussi par le streaming (puis recollage) pour éviter
+les coupures de passerelle sur les longues générations (cf. Cloudflare 524).
+"""
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+from typing import AsyncIterator
+
+import httpx
+
+from app.domain.models import ChatMessage
+from app.domain.ports import LLMGenerationTimeout, LLMProviderError
+
+logger = logging.getLogger(__name__)
+
+# Délai max pour le PREMIER token de contenu. Un modèle « en file d'attente »
+# n'envoie que des keep-alive (aucun contenu) → on échoue vite et clairement au
+# lieu de pendre. Le timeout réseau d'httpx ne suffit pas : des keep-alive font
+# « arriver des octets » et empêchent son read-timeout de se déclencher.
+_FIRST_TOKEN_TIMEOUT_SECONDS = 120.0
+
+
+class BaseOpenAICompatibleAdapter:
+    """Base des adapters clients d'une API OpenAI-compatible (chat/completions SSE).
+
+    Satisfait par duck typing les ports LLMProvider et LLMChatProvider. Les
+    sous-classes définissent : ``_provider_label``, ``_api_url``,
+    ``_supports_json_object`` (mode JSON natif), et surchargent au besoin
+    ``_headers`` / ``_error_for_status`` / les messages de timeout.
+    """
+
+    # Surchargés par les sous-classes.
+    _provider_label: str = "LLM"
+    _api_url: str = ""
+    _supports_json_object: bool = False
+
+    def __init__(self, api_key: str, model: str, timeout: int) -> None:
+        self._api_key = api_key
+        self._model = model
+        self._timeout = timeout
+
+    # --- Spécificités surchargeables ----------------------------------------
+
+    def _headers(self) -> dict[str, str]:
+        return {
+            "Authorization": f"Bearer {self._api_key}",
+            "Content-Type": "application/json",
+        }
+
+    def _first_token_timeout_message(self) -> str:
+        return (
+            f"Erreur {self._provider_label} : aucun contenu produit en "
+            f"{int(_FIRST_TOKEN_TIMEOUT_SECONDS)}s — le modèle est probablement en "
+            "file d'attente / saturé. Réessayez plus tard ou choisissez un autre modèle."
+        )
+
+    def _generation_timeout_message(self) -> str:
+        return (
+            f"Erreur {self._provider_label} : génération non terminée en {self._timeout}s. "
+            "Réduisez la taille des morceaux d'import, augmentez le timeout, ou changez de modèle."
+        )
+
+    def _error_for_status(self, status_code: int, detail: str) -> LLMProviderError:
+        """Erreur de domaine pour une réponse HTTP >= 400 (détail déjà lu)."""
+        return LLMProviderError(
+            f"Erreur {self._provider_label} (HTTP {status_code})"
+            + (f" : {detail[:500]}" if detail else "")
+        )
+
+    # --- API publique (ports) -----------------------------------------------
+
+    async def generate(
+        self,
+        prompt: str,
+        *,
+        output_format: str | None = None,
+        temperature: float | None = None,
+    ) -> str:
+        """One-shot via streaming (puis recollage), avec garde-fous au temps écoulé."""
+        return await self._collect_with_timeouts(
+            [ChatMessage(role="user", content=prompt)], temperature, output_format
+        )
+
+    async def stream_chat(
+        self,
+        messages: list[ChatMessage],
+        *,
+        system_prompt: str | None = None,
+        temperature: float | None = None,
+    ) -> AsyncIterator[str]:
+        async for token in self._stream(messages, system_prompt, temperature):
+            yield token
+
+    # --- Mécanique partagée -------------------------------------------------
+
+    async def _collect_with_timeouts(
+        self,
+        messages: list[ChatMessage],
+        temperature: float | None,
+        output_format: str | None,
+    ) -> str:
+        """Collecte le stream avec DEUX garde-fous au temps écoulé :
+        - 1er token borné (`_FIRST_TOKEN_TIMEOUT_SECONDS`) : détecte un modèle bloqué
+          en file d'attente (que des keep-alive, aucun contenu) → échec rapide ;
+        - ceiling global (`self._timeout`) : génération qui ne se termine jamais.
+        """
+        async def _collect() -> str:
+            chunks: list[str] = []
+            agen = self._stream(messages, None, temperature, output_format)
+            try:
+                while True:
+                    # Borne SEULEMENT l'attente du 1er token ; ensuite on laisse
+                    # générer (le ceiling global couvre le reste).
+                    first = _FIRST_TOKEN_TIMEOUT_SECONDS if not chunks else None
+                    try:
+                        token = await asyncio.wait_for(agen.__anext__(), timeout=first)
+                    except StopAsyncIteration:
+                        break
+                    except asyncio.TimeoutError:
+                        raise LLMProviderError(self._first_token_timeout_message())
+                    chunks.append(token)
+            finally:
+                await agen.aclose()
+            return "".join(chunks)
+
+        try:
+            return await asyncio.wait_for(_collect(), timeout=self._timeout)
+        except asyncio.TimeoutError as exc:
+            raise LLMGenerationTimeout(self._generation_timeout_message()) from exc
+
+    def _build_body(
+        self,
+        messages: list[ChatMessage],
+        system_prompt: str | None,
+        temperature: float | None,
+        output_format: str | None,
+    ) -> dict[str, object]:
+        payload_messages: list[dict[str, str]] = []
+        if system_prompt:
+            payload_messages.append({"role": "system", "content": system_prompt})
+        for m in messages:
+            payload_messages.append({"role": m.role, "content": m.content})
+
+        body: dict[str, object] = {
+            "model": self._model,
+            "messages": payload_messages,
+            "stream": True,
+        }
+        if temperature is not None:
+            body["temperature"] = temperature
+        # Mode JSON natif : supprime les fences ```json et le JSON invalide (retours
+        # à la ligne bruts), principale cause de morceaux d'import ignorés. Un SCHÉMA
+        # (dict) est traduit en json_object — suffisant, les grands modèles cloud
+        # respectent la structure demandée par le prompt. Désactivé pour les
+        # providers/modèles gratuits qui ne le supportent pas (réponse vide).
+        if self._supports_json_object and output_format is not None:
+            body["response_format"] = {"type": "json_object"}
+        return body
+
+    async def _stream(
+        self,
+        messages: list[ChatMessage],
+        system_prompt: str | None,
+        temperature: float | None,
+        output_format: str | None = None,
+    ) -> AsyncIterator[str]:
+        body = self._build_body(messages, system_prompt, temperature, output_format)
+        async with httpx.AsyncClient(timeout=self._timeout) as client:
+            try:
+                async with client.stream(
+                    "POST", self._api_url, headers=self._headers(), json=body
+                ) as response:
+                    if response.status_code >= 400:
+                        # En streaming le corps n'est pas lu automatiquement : on le
+                        # lit pour exposer le détail du provider (le 429 précise le
+                        # type de quota, le 401 la clé invalide…), sinon on n'a que
+                        # le code HTTP nu et le diagnostic est impossible.
+                        detail = (await response.aread()).decode("utf-8", "replace").strip()
+                        raise self._error_for_status(response.status_code, detail)
+                    async for token in self._parse_sse(response):
+                        yield token
+            except httpx.HTTPError as exc:
+                raise LLMProviderError(self._format_http_error(exc)) from exc
+
+    @staticmethod
+    async def _parse_sse(response: httpx.Response) -> AsyncIterator[str]:
+        """SSE OpenAI : lignes `data: {json}`, fin sur `data: [DONE]`."""
+        async for line in response.aiter_lines():
+            if not line or not line.startswith("data:"):
+                continue  # lignes vides ou commentaires keep-alive (`: ...`)
+            data = line[len("data:"):].strip()
+            if data == "[DONE]":
+                return
+            try:
+                obj = json.loads(data)
+            except json.JSONDecodeError:
+                continue
+            choices = obj.get("choices")
+            if not choices:
+                continue
+            delta = choices[0].get("delta") or {}
+            content = delta.get("content")
+            if content:
+                yield content
+
+    def _format_http_error(self, exc: httpx.HTTPError) -> str:
+        """Message lisible (timeout, quota 429, crédits 402, modèle inconnu…)."""
+        if isinstance(exc, httpx.TimeoutException):
+            return (
+                f"Erreur {self._provider_label} : délai dépassé (timeout {self._timeout}s). "
+                "Le modèle a mis trop de temps — réduis la taille des morceaux d'import ou "
+                "augmente le timeout."
+            )
+        detail = str(exc) or exc.__class__.__name__
+        return f"Erreur {self._provider_label} ({exc.__class__.__name__}) : {detail}"

brain/app/infrastructure/gemini_adapter.py

@@ -1,194 +1,66 @@
 """Adapter Google Gemini — implémente les ports LLMProvider / LLMChatProvider.
 
 Gemini expose un endpoint COMPATIBLE OpenAI
-(POST {base}/openai/chat/completions, SSE), donc cet adapter est un client
-"OpenAI-compatible" — même structure que les adapters OpenRouter / Mistral.
+(POST {base}/openai/chat/completions, SSE) : client "OpenAI-compatible" qui hérite
+de BaseOpenAICompatibleAdapter et ne fournit que ses spécificités (dont un message
+dédié quand Google refuse la clé en 401/403).
 
 Tier GRATUIT : clé API sur aistudio.google.com (sans CB). Atout majeur pour
 l'extraction de PDF : un CONTEXTE de ~1M tokens → un livre entier tient en 1-2
-appels, donc quasi aucun morceau perdu et peu de requêtes (limites jamais
-atteintes). Modèle conseillé : `gemini-2.0-flash` (rapide, gros contexte, fidèle).
+appels. Modèle conseillé : `gemini-2.0-flash` (rapide, gros contexte, fidèle).
 """
 from __future__ import annotations
 
-import asyncio
-import json
-import logging
-from typing import AsyncIterator
-
-import httpx
-
 from app.core.config import Settings
-from app.domain.models import ChatMessage
-from app.domain.ports import LLMGenerationTimeout, LLMProviderError
-
-logger = logging.getLogger(__name__)
-
-_API_URL = "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions"
-
-# Délai max pour le PREMIER token de contenu (échec rapide si le modèle ne produit
-# rien). Gemini répond vite ; 120s est large.
-_FIRST_TOKEN_TIMEOUT_SECONDS = 120.0
+from app.domain.ports import LLMProviderError
+from app.infrastructure.base_openai_adapter import (
+    _FIRST_TOKEN_TIMEOUT_SECONDS,
+    BaseOpenAICompatibleAdapter,
+)
 
 
-class GeminiLLMProvider:
+class GeminiLLMProvider(BaseOpenAICompatibleAdapter):
     """Adapter Gemini (OpenAI-compatible) — satisfait LLMProvider et LLMChatProvider."""
 
+    _provider_label = "Gemini"
+    _api_url = "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions"
+    _supports_json_object = True
+
     def __init__(self, settings: Settings) -> None:
         if not settings.gemini_api_key:
             raise LLMProviderError(
                 "Clé API Gemini manquante. Configure-la depuis l'écran Paramètres "
                 "(clé gratuite sur aistudio.google.com)."
             )
-        self._api_key = settings.gemini_api_key
-        self._model = settings.gemini_model
-        self._timeout = settings.llm_timeout_seconds
-
-    def _headers(self) -> dict[str, str]:
-        return {
-            "Authorization": f"Bearer {self._api_key}",
-            "Content-Type": "application/json",
-            "Accept": "application/json",
-        }
-
-    async def generate(
-        self,
-        prompt: str,
-        *,
-        output_format: str | None = None,
-        temperature: float | None = None,
-    ) -> str:
-        """One-shot via streaming (puis recollage), avec garde-fous au temps écoulé."""
-        return await self._collect_with_timeouts(
-            [ChatMessage(role="user", content=prompt)], temperature, output_format
+        super().__init__(
+            settings.gemini_api_key, settings.gemini_model, settings.llm_timeout_seconds
         )
 
-    async def _collect_with_timeouts(
-        self,
-        messages: list[ChatMessage],
-        temperature: float | None,
-        output_format: str | None,
-    ) -> str:
-        """Collecte le stream avec deux garde-fous : 1er token borné (échec rapide
-        si rien ne sort) + ceiling global `self._timeout`."""
-        async def _collect() -> str:
-            chunks: list[str] = []
-            agen = self._stream(messages, None, temperature, output_format)
-            try:
-                while True:
-                    first = _FIRST_TOKEN_TIMEOUT_SECONDS if not chunks else None
-                    try:
-                        token = await asyncio.wait_for(agen.__anext__(), timeout=first)
-                    except StopAsyncIteration:
-                        break
-                    except asyncio.TimeoutError:
-                        raise LLMProviderError(
-                            f"Erreur Gemini : aucun contenu produit en "
-                            f"{int(_FIRST_TOKEN_TIMEOUT_SECONDS)}s. Réessayez ou vérifiez "
-                            "votre quota gratuit."
-                        )
-                    chunks.append(token)
-            finally:
-                await agen.aclose()
-            return "".join(chunks)
+    def _headers(self) -> dict[str, str]:
+        return {**super()._headers(), "Accept": "application/json"}
 
-        try:
-            return await asyncio.wait_for(_collect(), timeout=self._timeout)
-        except asyncio.TimeoutError as exc:
-            raise LLMGenerationTimeout(
-                f"Erreur Gemini : génération non terminée en {self._timeout}s. Réduisez la "
-                "taille des morceaux d'import ou augmentez le timeout."
-            ) from exc
+    def _first_token_timeout_message(self) -> str:
+        return (
+            f"Erreur Gemini : aucun contenu produit en "
+            f"{int(_FIRST_TOKEN_TIMEOUT_SECONDS)}s. Réessayez ou vérifiez "
+            "votre quota gratuit."
+        )
 
-    async def stream_chat(
-        self,
-        messages: list[ChatMessage],
-        *,
-        system_prompt: str | None = None,
-        temperature: float | None = None,
-    ) -> AsyncIterator[str]:
-        async for token in self._stream(messages, system_prompt, temperature):
-            yield token
+    def _generation_timeout_message(self) -> str:
+        return (
+            f"Erreur Gemini : génération non terminée en {self._timeout}s. Réduisez la "
+            "taille des morceaux d'import ou augmentez le timeout."
+        )
 
-    async def _stream(
-        self,
-        messages: list[ChatMessage],
-        system_prompt: str | None,
-        temperature: float | None,
-        output_format: str | None = None,
-    ) -> AsyncIterator[str]:
-        payload_messages: list[dict[str, str]] = []
-        if system_prompt:
-            payload_messages.append({"role": "system", "content": system_prompt})
-        for m in messages:
-            payload_messages.append({"role": m.role, "content": m.content})
-
-        body: dict[str, object] = {
-            "model": self._model,
-            "messages": payload_messages,
-            "stream": True,
-        }
-        if temperature is not None:
-            body["temperature"] = temperature
-        # Mode JSON natif (supporté par l'endpoint OpenAI-compatible de Gemini) :
-        # supprime fences ```json et JSON invalide, principale cause de morceaux
-        # ignorés. Un SCHÉMA (dict) est traduit en json_object : suffisant, les
-        # grands modèles cloud respectent la structure demandée par le prompt.
-        if output_format is not None:
-            body["response_format"] = {"type": "json_object"}
-
-        async with httpx.AsyncClient(timeout=self._timeout) as client:
-            try:
-                async with client.stream(
-                    "POST", _API_URL, headers=self._headers(), json=body
-                ) as response:
-                    if response.status_code >= 400:
-                        detail = (await response.aread()).decode("utf-8", "replace").strip()
-                        # 401/403 = clé rejetée par GOOGLE (pas un problème LoreMind) :
-                        # message actionnable plutôt que le JSON brut de l'API.
-                        if response.status_code in (401, 403):
-                            raise LLMProviderError(
-                                "Erreur Gemini : clé API refusée par Google "
-                                f"(HTTP {response.status_code}). Vérifiez que la clé vient bien "
-                                "de aistudio.google.com (« Get API key ») et qu'elle n'a pas de "
-                                "restrictions (API ou adresse IP) dans la Google Cloud Console. "
-                                f"Détail : {detail[:300]}"
-                            )
-                        raise LLMProviderError(
-                            f"Erreur Gemini (HTTP {response.status_code})"
-                            + (f" : {detail[:500]}" if detail else "")
-                        )
-                    async for token in self._parse_sse(response):
-                        yield token
-            except httpx.HTTPError as exc:
-                raise LLMProviderError(self._format_http_error(exc)) from exc
-
-    @staticmethod
-    async def _parse_sse(response: httpx.Response) -> AsyncIterator[str]:
-        """SSE OpenAI : lignes `data: {json}`, fin sur `data: [DONE]`."""
-        async for line in response.aiter_lines():
-            if not line or not line.startswith("data:"):
-                continue
-            data = line[len("data:"):].strip()
-            if data == "[DONE]":
-                return
-            try:
-                obj = json.loads(data)
-            except json.JSONDecodeError:
-                continue
-            choices = obj.get("choices")
-            if not choices:
-                continue
-            delta = choices[0].get("delta") or {}
-            content = delta.get("content")
-            if content:
-                yield content
-
-    def _format_http_error(self, exc: httpx.HTTPError) -> str:
-        if isinstance(exc, httpx.TimeoutException):
-            return (
-                f"Erreur Gemini : délai dépassé (timeout {self._timeout}s). Le modèle a "
-                "mis trop de temps — réduis la taille des morceaux d'import ou augmente le timeout."
+    def _error_for_status(self, status_code: int, detail: str) -> LLMProviderError:
+        # 401/403 = clé rejetée par GOOGLE (pas un problème LoreMind) : message
+        # actionnable plutôt que le JSON brut de l'API.
+        if status_code in (401, 403):
+            return LLMProviderError(
+                "Erreur Gemini : clé API refusée par Google "
+                f"(HTTP {status_code}). Vérifiez que la clé vient bien "
+                "de aistudio.google.com (« Get API key ») et qu'elle n'a pas de "
+                "restrictions (API ou adresse IP) dans la Google Cloud Console. "
+                f"Détail : {detail[:300]}"
             )
-        detail = str(exc) or exc.__class__.__name__
-        return f"Erreur Gemini ({exc.__class__.__name__}) : {detail}"
+        return super()._error_for_status(status_code, detail)

brain/app/infrastructure/mistral_adapter.py

@@ -1,195 +1,48 @@
 """Adapter Mistral — implémente les ports LLMProvider / LLMChatProvider.
 
 Mistral (La Plateforme) expose l'API OpenAI standard (POST {base}/chat/completions,
-SSE), donc cet adapter est un client "OpenAI-compatible" — même structure que
-l'adapter OpenRouter. Le `generate` one-shot passe par le streaming (puis
-recollage) avec un timeout au temps écoulé pour ne jamais pendre à l'infini.
+SSE) : client "OpenAI-compatible" qui hérite de BaseOpenAICompatibleAdapter et ne
+fournit que ses spécificités.
 
 Tier GRATUIT : compte sur console.mistral.ai (tier « Experiment »), clé API à
 coller dans l'écran Paramètres. Modèles conseillés pour l'extraction : un grand
 contexte fidèle comme `mistral-large-latest` (128k) ou `mistral-small-latest`.
+
+Mode JSON natif : TOUS les modèles Mistral le supportent (`_supports_json_object`).
 """
 from __future__ import annotations
 
-import asyncio
-import json
-import logging
-from typing import AsyncIterator
-
-import httpx
-
 from app.core.config import Settings
-from app.domain.models import ChatMessage
-from app.domain.ports import LLMGenerationTimeout, LLMProviderError
-
-logger = logging.getLogger(__name__)
-
-_API_URL = "https://api.mistral.ai/v1/chat/completions"
-
-# Délai max pour le PREMIER token de contenu (échec rapide si le modèle est en file
-# d'attente et n'envoie que des keep-alive). Généreux car la file d'un tier gratuit
-# peut être longue.
-_FIRST_TOKEN_TIMEOUT_SECONDS = 120.0
+from app.domain.ports import LLMProviderError
+from app.infrastructure.base_openai_adapter import (
+    _FIRST_TOKEN_TIMEOUT_SECONDS,
+    BaseOpenAICompatibleAdapter,
+)
 
 
-class MistralLLMProvider:
+class MistralLLMProvider(BaseOpenAICompatibleAdapter):
     """Adapter Mistral (OpenAI-compatible) — satisfait LLMProvider et LLMChatProvider."""
 
+    _provider_label = "Mistral"
+    _api_url = "https://api.mistral.ai/v1/chat/completions"
+    _supports_json_object = True
+
     def __init__(self, settings: Settings) -> None:
         if not settings.mistral_api_key:
             raise LLMProviderError(
                 "Clé API Mistral manquante. Configure-la depuis l'écran Paramètres."
             )
-        self._api_key = settings.mistral_api_key
-        self._model = settings.mistral_model
-        self._timeout = settings.llm_timeout_seconds
-
-    def _headers(self) -> dict[str, str]:
-        return {
-            "Authorization": f"Bearer {self._api_key}",
-            "Content-Type": "application/json",
-            "Accept": "application/json",
-        }
-
-    async def generate(
-        self,
-        prompt: str,
-        *,
-        output_format: str | None = None,
-        temperature: float | None = None,
-    ) -> str:
-        """One-shot via streaming (puis recollage) pour robustesse sur longues sorties.
-
-        Timeout au TEMPS ÉCOULÉ (asyncio) en plus du timeout réseau d'httpx :
-        si le provider envoyait des keep-alive sans contenu, l'appel pendrait à
-        l'infini. Ici on coupe net après `self._timeout` secondes.
-        """
-        return await self._collect_with_timeouts(
-            [ChatMessage(role="user", content=prompt)], temperature, output_format
+        super().__init__(
+            settings.mistral_api_key, settings.mistral_model, settings.llm_timeout_seconds
         )
 
-    async def _collect_with_timeouts(
-        self,
-        messages: list[ChatMessage],
-        temperature: float | None,
-        output_format: str | None,
-    ) -> str:
-        """Collecte le stream avec deux garde-fous au temps écoulé : 1er token borné
-        (file d'attente → échec rapide) + ceiling global `self._timeout`."""
-        async def _collect() -> str:
-            chunks: list[str] = []
-            agen = self._stream(messages, None, temperature, output_format)
-            try:
-                while True:
-                    first = _FIRST_TOKEN_TIMEOUT_SECONDS if not chunks else None
-                    try:
-                        token = await asyncio.wait_for(agen.__anext__(), timeout=first)
-                    except StopAsyncIteration:
-                        break
-                    except asyncio.TimeoutError:
-                        raise LLMProviderError(
-                            f"Erreur Mistral : aucun contenu produit en "
-                            f"{int(_FIRST_TOKEN_TIMEOUT_SECONDS)}s — le modèle est probablement "
-                            "en file d'attente (tier gratuit, 2 req/min). Réessayez plus tard ou "
-                            "choisissez un modèle plus disponible."
-                        )
-                    chunks.append(token)
-            finally:
-                await agen.aclose()
-            return "".join(chunks)
+    def _headers(self) -> dict[str, str]:
+        return {**super()._headers(), "Accept": "application/json"}
 
-        try:
-            return await asyncio.wait_for(_collect(), timeout=self._timeout)
-        except asyncio.TimeoutError as exc:
-            raise LLMGenerationTimeout(
-                f"Erreur Mistral : génération non terminée en {self._timeout}s. Réduisez la "
-                "taille des morceaux d'import, augmentez le timeout, ou changez de modèle."
-            ) from exc
-
-    async def stream_chat(
-        self,
-        messages: list[ChatMessage],
-        *,
-        system_prompt: str | None = None,
-        temperature: float | None = None,
-    ) -> AsyncIterator[str]:
-        async for token in self._stream(messages, system_prompt, temperature):
-            yield token
-
-    async def _stream(
-        self,
-        messages: list[ChatMessage],
-        system_prompt: str | None,
-        temperature: float | None,
-        output_format: str | None = None,
-    ) -> AsyncIterator[str]:
-        payload_messages: list[dict[str, str]] = []
-        if system_prompt:
-            payload_messages.append({"role": "system", "content": system_prompt})
-        for m in messages:
-            payload_messages.append({"role": m.role, "content": m.content})
-
-        body: dict[str, object] = {
-            "model": self._model,
-            "messages": payload_messages,
-            "stream": True,
-        }
-        if temperature is not None:
-            body["temperature"] = temperature
-        # Mode JSON natif : TOUS les modèles Mistral le supportent → plus de fences
-        # ```json ni de JSON invalide (retours à la ligne bruts dans les chaînes),
-        # principale cause de morceaux d'import ignorés. Un SCHÉMA (dict) est
-        # traduit en json_object : suffisant ici, les grands modèles cloud
-        # respectent la structure demandée par le prompt.
-        if output_format is not None:
-            body["response_format"] = {"type": "json_object"}
-
-        async with httpx.AsyncClient(timeout=self._timeout) as client:
-            try:
-                async with client.stream(
-                    "POST", _API_URL, headers=self._headers(), json=body
-                ) as response:
-                    if response.status_code >= 400:
-                        # En streaming le corps n'est pas lu automatiquement : on le
-                        # lit pour exposer le détail de Mistral (modèle inconnu, clé
-                        # invalide 401, quota 429…), sinon on n'a que le code HTTP nu.
-                        detail = (await response.aread()).decode("utf-8", "replace").strip()
-                        raise LLMProviderError(
-                            f"Erreur Mistral (HTTP {response.status_code})"
-                            + (f" : {detail[:500]}" if detail else "")
-                        )
-                    async for token in self._parse_sse(response):
-                        yield token
-            except httpx.HTTPError as exc:
-                raise LLMProviderError(self._format_http_error(exc)) from exc
-
-    @staticmethod
-    async def _parse_sse(response: httpx.Response) -> AsyncIterator[str]:
-        """SSE OpenAI : lignes `data: {json}`, fin sur `data: [DONE]`."""
-        async for line in response.aiter_lines():
-            if not line or not line.startswith("data:"):
-                continue  # lignes vides ou keep-alive (`: ...`)
-            data = line[len("data:"):].strip()
-            if data == "[DONE]":
-                return
-            try:
-                obj = json.loads(data)
-            except json.JSONDecodeError:
-                continue
-            choices = obj.get("choices")
-            if not choices:
-                continue
-            delta = choices[0].get("delta") or {}
-            content = delta.get("content")
-            if content:
-                yield content
-
-    def _format_http_error(self, exc: httpx.HTTPError) -> str:
-        """Message lisible (timeout, quota 429, clé invalide 401, modèle inconnu…)."""
-        if isinstance(exc, httpx.TimeoutException):
-            return (
-                f"Erreur Mistral : délai dépassé (timeout {self._timeout}s). Le modèle a "
-                "mis trop de temps — réduis la taille des morceaux d'import ou augmente le timeout."
-            )
-        detail = str(exc) or exc.__class__.__name__
-        return f"Erreur Mistral ({exc.__class__.__name__}) : {detail}"
+    def _first_token_timeout_message(self) -> str:
+        return (
+            f"Erreur Mistral : aucun contenu produit en "
+            f"{int(_FIRST_TOKEN_TIMEOUT_SECONDS)}s — le modèle est probablement "
+            "en file d'attente (tier gratuit, 2 req/min). Réessayez plus tard ou "
+            "choisissez un modèle plus disponible."
+        )

brain/app/infrastructure/openrouter_adapter.py

@@ -1,205 +1,57 @@
 """Adapter OpenRouter — implémente les ports LLMProvider / LLMChatProvider.
 
 OpenRouter expose l'API OpenAI standard (POST {base}/chat/completions, SSE), donc
-cet adapter est en réalité un client "OpenAI-compatible". Le `generate` one-shot
-passe lui aussi par le streaming (puis recollage) pour éviter les coupures de
-passerelle sur les longues générations (cf. 1min.ai / Cloudflare 524).
+cet adapter est un client "OpenAI-compatible" : il hérite de toute la mécanique de
+BaseOpenAICompatibleAdapter et ne fournit que ses spécificités (URL, en-têtes
+d'attribution, messages, lecture de config).
 
 Modèles GRATUITS : utiliser un id finissant par `:free` (ex.
 `meta-llama/llama-3.3-70b-instruct:free`) ou le routeur `openrouter/free` (défaut)
 qui choisit automatiquement un modèle gratuit — aucun crédit consommé.
+
+NB : on n'impose PAS `response_format=json_object` (`_supports_json_object=False`).
+Beaucoup de modèles/providers GRATUITS ne le supportent pas et renvoient une
+réponse VIDE. On laisse le modèle répondre librement ; l'extraction JSON en aval
+(load_json_object + nettoyage du raisonnement) récupère le JSON dans la prose.
 """
 from __future__ import annotations
 
-import asyncio
-import json
-import logging
-from typing import AsyncIterator
-
-import httpx
-
 from app.core.config import Settings
-from app.domain.models import ChatMessage
-
-logger = logging.getLogger(__name__)
-from app.domain.ports import LLMGenerationTimeout, LLMProviderError
-
-_API_URL = "https://openrouter.ai/api/v1/chat/completions"
-
-# Délai max pour le PREMIER token de contenu. Un modèle gratuit "en file d'attente"
-# n'envoie que des keep-alive (aucun contenu) → on échoue vite et clairement au lieu
-# de pendre. Généreux (2 min) car la file d'attente d'un tier gratuit peut être longue.
-_FIRST_TOKEN_TIMEOUT_SECONDS = 120.0
+from app.domain.ports import LLMProviderError
+from app.infrastructure.base_openai_adapter import (
+    _FIRST_TOKEN_TIMEOUT_SECONDS,
+    BaseOpenAICompatibleAdapter,
+)
 
 
-class OpenRouterLLMProvider:
+class OpenRouterLLMProvider(BaseOpenAICompatibleAdapter):
     """Adapter OpenRouter (OpenAI-compatible) — satisfait LLMProvider et LLMChatProvider."""
 
+    _provider_label = "OpenRouter"
+    _api_url = "https://openrouter.ai/api/v1/chat/completions"
+    _supports_json_object = False
+
     def __init__(self, settings: Settings) -> None:
         if not settings.openrouter_api_key:
             raise LLMProviderError(
                 "Clé API OpenRouter manquante. Configure-la depuis l'écran Paramètres."
             )
-        self._api_key = settings.openrouter_api_key
-        self._model = settings.openrouter_model
-        self._timeout = settings.llm_timeout_seconds
+        super().__init__(
+            settings.openrouter_api_key, settings.openrouter_model, settings.llm_timeout_seconds
+        )
 
     def _headers(self) -> dict[str, str]:
         return {
-            "Authorization": f"Bearer {self._api_key}",
-            "Content-Type": "application/json",
+            **super()._headers(),
             # Attribution facultative (classement OpenRouter) — sans impact fonctionnel.
             "HTTP-Referer": "https://loremind.app",
             "X-Title": "LoreMind",
         }
 
-    async def generate(
-        self,
-        prompt: str,
-        *,
-        output_format: str | None = None,
-        temperature: float | None = None,
-    ) -> str:
-        """One-shot via streaming (puis recollage) pour robustesse sur longues sorties.
-
-        Timeout au TEMPS ÉCOULÉ (asyncio) en plus du timeout réseau d'httpx : un
-        modèle gratuit saturé/en file d'attente envoie des keep-alive (`: OPENROUTER
-        PROCESSING`) mais AUCUN contenu → httpx ne déclenche jamais son read-timeout
-        (des octets arrivent) et l'appel pendrait à l'infini. Ici on coupe net après
-        `self._timeout` secondes, quoi qu'il arrive.
-        """
-        return await self._collect_with_timeouts(
-            [ChatMessage(role="user", content=prompt)], temperature, output_format, "OpenRouter"
+    def _first_token_timeout_message(self) -> str:
+        return (
+            f"Erreur OpenRouter : aucun contenu produit en "
+            f"{int(_FIRST_TOKEN_TIMEOUT_SECONDS)}s — le modèle gratuit est "
+            "probablement en file d'attente / saturé. Réessayez plus tard ou "
+            "choisissez un autre modèle (1min.ai, ou payant)."
         )
-
-    async def _collect_with_timeouts(
-        self,
-        messages: list[ChatMessage],
-        temperature: float | None,
-        output_format: str | None,
-        provider: str,
-    ) -> str:
-        """Collecte le stream avec DEUX garde-fous au temps écoulé :
-        - 1er token borné (`_FIRST_TOKEN_TIMEOUT_SECONDS`) : détecte un modèle bloqué
-          en file d'attente (que des keep-alive, aucun contenu) → échec rapide ;
-        - ceiling global (`self._timeout`) : génération qui ne se termine jamais.
-        Le timeout réseau d'httpx ne suffit pas : des keep-alive font 'arriver des
-        octets' et empêchent son read-timeout de se déclencher.
-        """
-        async def _collect() -> str:
-            chunks: list[str] = []
-            agen = self._stream(messages, None, temperature, output_format)
-            try:
-                while True:
-                    # Borne SEULEMENT l'attente du 1er token (file d'attente) ; ensuite
-                    # on laisse générer (le ceiling global couvre le reste).
-                    first = _FIRST_TOKEN_TIMEOUT_SECONDS if not chunks else None
-                    try:
-                        token = await asyncio.wait_for(agen.__anext__(), timeout=first)
-                    except StopAsyncIteration:
-                        break
-                    except asyncio.TimeoutError:
-                        raise LLMProviderError(
-                            f"Erreur {provider} : aucun contenu produit en "
-                            f"{int(_FIRST_TOKEN_TIMEOUT_SECONDS)}s — le modèle gratuit est "
-                            "probablement en file d'attente / saturé. Réessayez plus tard ou "
-                            "choisissez un autre modèle (1min.ai, ou payant)."
-                        )
-                    chunks.append(token)
-            finally:
-                await agen.aclose()
-            return "".join(chunks)
-
-        try:
-            return await asyncio.wait_for(_collect(), timeout=self._timeout)
-        except asyncio.TimeoutError as exc:
-            raise LLMGenerationTimeout(
-                f"Erreur {provider} : génération non terminée en {self._timeout}s. Réduisez la "
-                "taille des morceaux d'import, augmentez le timeout, ou changez de modèle."
-            ) from exc
-
-    async def stream_chat(
-        self,
-        messages: list[ChatMessage],
-        *,
-        system_prompt: str | None = None,
-        temperature: float | None = None,
-    ) -> AsyncIterator[str]:
-        async for token in self._stream(messages, system_prompt, temperature):
-            yield token
-
-    async def _stream(
-        self,
-        messages: list[ChatMessage],
-        system_prompt: str | None,
-        temperature: float | None,
-        output_format: str | None = None,
-    ) -> AsyncIterator[str]:
-        payload_messages: list[dict[str, str]] = []
-        if system_prompt:
-            payload_messages.append({"role": "system", "content": system_prompt})
-        for m in messages:
-            payload_messages.append({"role": m.role, "content": m.content})
-
-        body: dict[str, object] = {
-            "model": self._model,
-            "messages": payload_messages,
-            "stream": True,
-        }
-        if temperature is not None:
-            body["temperature"] = temperature
-        # NB : on n'impose PAS `response_format=json_object`. Beaucoup de modèles/
-        # providers GRATUITS ne le supportent pas et renvoient une réponse VIDE.
-        # On laisse le modèle répondre librement ; l'extraction JSON en aval
-        # (load_json_object + nettoyage du raisonnement) récupère le JSON dans la prose.
-
-        async with httpx.AsyncClient(timeout=self._timeout) as client:
-            try:
-                async with client.stream(
-                    "POST", _API_URL, headers=self._headers(), json=body
-                ) as response:
-                    if response.status_code >= 400:
-                        # En streaming, le corps n'est pas lu automatiquement : on le
-                        # lit pour exposer le détail d'OpenRouter (ex. le 429 précise
-                        # "free-models-per-day" vs "per-minute"), sinon on n'a que le
-                        # code HTTP nu et le diagnostic est impossible.
-                        detail = (await response.aread()).decode("utf-8", "replace").strip()
-                        raise LLMProviderError(
-                            f"Erreur OpenRouter (HTTP {response.status_code})"
-                            + (f" : {detail[:500]}" if detail else "")
-                        )
-                    async for token in self._parse_sse(response):
-                        yield token
-            except httpx.HTTPError as exc:
-                raise LLMProviderError(self._format_http_error(exc)) from exc
-
-    @staticmethod
-    async def _parse_sse(response: httpx.Response) -> AsyncIterator[str]:
-        """SSE OpenAI : lignes `data: {json}`, fin sur `data: [DONE]`."""
-        async for line in response.aiter_lines():
-            if not line or not line.startswith("data:"):
-                continue  # lignes vides ou commentaires keep-alive (`: ...`)
-            data = line[len("data:"):].strip()
-            if data == "[DONE]":
-                return
-            try:
-                obj = json.loads(data)
-            except json.JSONDecodeError:
-                continue
-            choices = obj.get("choices")
-            if not choices:
-                continue
-            delta = choices[0].get("delta") or {}
-            content = delta.get("content")
-            if content:
-                yield content
-
-    def _format_http_error(self, exc: httpx.HTTPError) -> str:
-        """Message lisible (timeout, quota 429, crédits 402, modèle inconnu…)."""
-        if isinstance(exc, httpx.TimeoutException):
-            return (
-                f"Erreur OpenRouter : délai dépassé (timeout {self._timeout}s). Le modèle a "
-                "mis trop de temps — réduis la taille des morceaux d'import ou augmente le timeout."
-            )
-        detail = str(exc) or exc.__class__.__name__
-        return f"Erreur OpenRouter ({exc.__class__.__name__}) : {detail}"

brain/app/main.py

@@ -26,7 +26,7 @@ from app.infrastructure.ollama_model_installer import ensure_ollama_embedding_mo
 app = FastAPI(
     title="LoreMind Brain",
     description="Backend IA pour la génération de contenu narratif.",
-    version="0.16.1",
+    version="0.16.2",
 )
 
 logger = logging.getLogger(__name__)

brain/pytest.ini

@@ -0,0 +1,7 @@
+[pytest]
+# Tests unitaires du brain. asyncio_mode=auto : les coroutines de test sont
+# exécutées sans décorateur @pytest.mark.asyncio explicite.
+asyncio_mode = auto
+# Ajoute la racine du brain au sys.path pour `import app...` sans installation.
+pythonpath = .
+testpaths = tests

brain/requirements-dev.txt

@@ -0,0 +1,11 @@
+# Dépendances de TEST uniquement (non embarquées dans l'image / le bundle desktop).
+# Installer avec : .venv/Scripts/python -m pip install -r requirements-dev.txt
+-r requirements.txt
+
+pytest>=8,<9
+pytest-asyncio>=0.24,<1
+# Mock du transport httpx (intercepte les appels aux API LLM dans les tests).
+respx>=0.21,<1
+# Couverture de tests (équivalent JaCoCo) : `pytest --cov=app --cov-report=html`
+# → rapport HTML dans htmlcov/. La CI ajoute --cov-fail-under pour le plancher.
+pytest-cov>=5,<7

brain/tests/test_adapt_campaign.py

@@ -0,0 +1,72 @@
+"""Tests du use case de conseils d'adaptation (app.application.adapt_campaign)."""
+from __future__ import annotations
+
+import pytest
+
+from app.application.adapt_campaign import AdaptCampaignUseCase
+from app.domain.models import ChatMessage, ExtractedDocument, ExtractedPage
+from app.domain.ports import PdfExtractionError
+
+
+class FakeExtractor:
+    def __init__(self, doc: ExtractedDocument) -> None:
+        self._doc = doc
+
+    def extract(self, pdf_bytes: bytes) -> ExtractedDocument:
+        return self._doc
+
+
+class FakeChatLLM:
+    def __init__(self, tokens: list[str]) -> None:
+        self._tokens = tokens
+        self.system_prompt: str | None = None
+        self.messages: list[ChatMessage] | None = None
+
+    async def stream_chat(self, messages, *, system_prompt=None, temperature=None):
+        self.messages = messages
+        self.system_prompt = system_prompt
+        for t in self._tokens:
+            yield t
+
+
+def _doc(text: str) -> ExtractedDocument:
+    return ExtractedDocument(pages=[ExtractedPage(index=0, text=text, used_ocr=False)])
+
+
+async def test_stream_yields_tokens_and_builds_context():
+    llm = FakeChatLLM(["con", "seil"])
+    uc = AdaptCampaignUseCase(llm, FakeExtractor(_doc("contenu du pdf")))
+    out = [t async for t in uc.stream(b"x", "mon brief de campagne",
+                                      [ChatMessage(role="user", content="aide")])]
+    assert out == ["con", "seil"]
+    assert "mon brief de campagne" in llm.system_prompt
+    assert "contenu du pdf" in llm.system_prompt
+
+
+async def test_stream_empty_pdf_text_raises():
+    uc = AdaptCampaignUseCase(FakeChatLLM([]), FakeExtractor(_doc("   ")))
+    with pytest.raises(PdfExtractionError):
+        [t async for t in uc.stream(b"x", "brief", [])]
+
+
+async def test_stream_injects_default_request_when_no_messages():
+    llm = FakeChatLLM(["ok"])
+    uc = AdaptCampaignUseCase(llm, FakeExtractor(_doc("texte du pdf")))
+    _ = [t async for t in uc.stream(b"x", "", [])]
+    assert llm.messages[0].role == "user"
+    assert "campagne" in llm.messages[0].content.lower()
+
+
+def test_fit_pdf_short_text_not_truncated():
+    uc = AdaptCampaignUseCase(None, None, max_input_tokens=10000)
+    text, truncated = uc._fit_pdf_to_budget("court texte", "brief")
+    assert truncated is False
+    assert text == "court texte"
+
+
+def test_fit_pdf_long_text_is_truncated():
+    uc = AdaptCampaignUseCase(None, None, max_input_tokens=2100)
+    long_text = "mot " * 5000
+    text, truncated = uc._fit_pdf_to_budget(long_text, "")
+    assert truncated is True
+    assert len(text) < len(long_text)

brain/tests/test_chat.py

@@ -0,0 +1,129 @@
+"""Tests de la construction du system prompt du chat (app.application.chat).
+
+Assertions par INCLUSION (présence des données/sections clés) plutôt que sur le
+texte exact des consignes : robuste aux retouches de formulation, tout en
+vérifiant que chaque contexte est bien injecté.
+"""
+from __future__ import annotations
+
+from app.application.chat import ChatUseCase
+from app.domain.models import (
+    ArcSummary,
+    CampaignStructuralContext,
+    ChapterSummary,
+    CharacterSummary,
+    ChatMessage,
+    GameSystemContext,
+    JournalEntrySummary,
+    LoreStructuralContext,
+    NarrativeEntityContext,
+    NpcSummary,
+    PageContext,
+    PageSummary,
+    SceneSummary,
+    SessionContext,
+)
+
+
+def _build(**kw) -> str:
+    return ChatUseCase(None).build_system_prompt(**kw)
+
+
+def _empty_lore() -> LoreStructuralContext:
+    return LoreStructuralContext(lore_name="L", lore_description=None, folders={}, tags=[])
+
+
+def test_base_prompt_without_context_is_non_empty():
+    assert len(_build()) > 0
+
+
+def test_lore_block_renders_pages_with_values_tags_and_links():
+    lore = LoreStructuralContext(
+        lore_name="Eldoria", lore_description="un monde sombre",
+        folders={"PNJ": [PageSummary(
+            title="Aragorn", template_name="Personnage",
+            values={"apparence": "grand et noble"}, tags=["héros"],
+            related_page_titles=["Gondor"])]},
+        tags=["dark-fantasy"])
+    p = _build(lore_context=lore)
+    assert "Eldoria" in p
+    assert "un monde sombre" in p
+    assert "Aragorn" in p
+    assert "apparence" in p and "grand et noble" in p
+    assert "héros" in p
+    assert "Gondor" in p
+
+
+def test_empty_lore_signals_vide():
+    assert "Lore vide" in _build(lore_context=_empty_lore())
+
+
+def test_page_context_block_lists_fields_and_empty_marker():
+    page = PageContext(title="Aragorn", template_name="Personnage",
+                       template_fields=["apparence", "histoire"],
+                       values={"apparence": "grand"})
+    p = _build(page_context=page)
+    assert "PAGE EN COURS" in p
+    assert "Aragorn" in p
+    assert "apparence" in p
+    assert "(vide)" in p  # 'histoire' sans valeur
+
+
+def test_campaign_block_with_arc_and_empty_pj_npc_and_no_lore_note():
+    camp = CampaignStructuralContext(
+        campaign_name="La Malédiction", campaign_description="horreur gothique",
+        arcs=[ArcSummary(name="Acte I", description="intro",
+                         chapters=[ChapterSummary(name="Ch1", description="",
+                                                  scenes=[SceneSummary(name="Sc1", description="")])])],
+        characters=[], npcs=[])
+    p = _build(campaign_context=camp)
+    assert "CAMPAGNE COURANTE" in p
+    assert "La Malédiction" in p
+    assert "Acte I" in p
+    assert "aucune fiche" in p       # pas de PJ
+    assert "aucun univers" in p      # pas de lore lié
+
+
+def test_campaign_with_characters_npcs_and_lore_present_note():
+    camp = CampaignStructuralContext(
+        campaign_name="C", campaign_description=None, arcs=[],
+        characters=[CharacterSummary(name="Tav", snippet="roublarde")],
+        npcs=[NpcSummary(name="Strahd", snippet="vampire de Barovia")])
+    p = _build(campaign_context=camp, lore_context=_empty_lore())
+    assert "Tav" in p and "roublarde" in p
+    assert "Strahd" in p and "vampire de Barovia" in p
+    assert "liée à l'univers" in p
+
+
+def test_game_system_narrative_and_session_sections_injected():
+    gs = GameSystemContext(system_name="Nimble", system_description=None,
+                           sections={"Combat": "règles de combat"})
+    narr = NarrativeEntityContext(entity_type="scene", title="L'auberge du Portail",
+                                  fields={"ambiance": "tendue"})
+    sess = SessionContext(
+        session_name="Séance 3", active=True, started_at=None,
+        entries=[JournalEntrySummary(type="EVENT", content="Le pont s'effondre", occurred_at=None)],
+        previous_events=[])
+    p = _build(lore_context=_empty_lore(), game_system_context=gs,
+               narrative_entity=narr, session_context=sess)
+    assert "Nimble" in p
+    assert "L'auberge du Portail" in p
+    assert "Séance 3" in p
+    assert "Le pont s'effondre" in p
+
+
+async def test_stream_passes_built_prompt_to_llm():
+    class FakeChatLLM:
+        def __init__(self) -> None:
+            self.system_prompt = None
+
+        async def stream_chat(self, messages, *, system_prompt=None, temperature=None):
+            self.system_prompt = system_prompt
+            yield "tok"
+
+    llm = FakeChatLLM()
+    out = [t async for t in ChatUseCase(llm).stream(
+        [ChatMessage(role="user", content="salut")],
+        lore_context=LoreStructuralContext("Eldoria", None, {}, []))]
+    assert out == ["tok"]
+    assert "Eldoria" in llm.system_prompt

brain/tests/test_chunking.py

@@ -0,0 +1,72 @@
+"""Tests du découpage de texte (app.application.chunking).
+
+Vérifie le découpage par paragraphes vers une cible de tokens, le découpage des
+paragraphes géants, le recouvrement (overlap), et le split_in_half du repli
+anti-troncature. tiktoken (cl100k_base) est déterministe → assertions stables.
+"""
+from __future__ import annotations
+
+from app.application.chunking import chunk_text, split_in_half
+
+
+def test_empty_text_returns_no_chunk():
+    assert chunk_text("") == []
+    assert chunk_text("   \n\n  ") == []
+
+
+def test_short_text_stays_single_chunk():
+    chunks = chunk_text("Paragraphe un.\n\nParagraphe deux.", target_tokens=1000)
+    assert len(chunks) == 1
+    assert "Paragraphe un." in chunks[0]
+    assert "Paragraphe deux." in chunks[0]
+
+
+def test_splits_into_several_chunks_when_exceeding_target():
+    paras = [f"Paragraphe numero {i} avec un peu de contenu." for i in range(20)]
+    full = "\n\n".join(paras)
+    chunks = chunk_text(full, target_tokens=20)
+    assert len(chunks) > 1
+    # Aucun paragraphe perdu : tous présents quelque part.
+    joined = "\n\n".join(chunks)
+    for p in paras:
+        assert p in joined
+
+
+def test_oversized_single_paragraph_is_split():
+    # Un seul paragraphe (aucun "\n\n") plus gros que la cible → plusieurs sous-blocs.
+    huge = "mot " * 500
+    chunks = chunk_text(huge, target_tokens=50)
+    assert len(chunks) > 1
+
+
+def test_overlap_repeats_content_without_losing_paragraphs():
+    paras = [f"Bloc {i} de texte distinct." for i in range(12)]
+    full = "\n\n".join(paras)
+    chunks = chunk_text(full, target_tokens=20, overlap_tokens=10)
+    assert len(chunks) > 1
+    joined = "\n\n".join(chunks)
+    for p in paras:
+        assert p in joined
+
+
+# --- split_in_half -------------------------------------------------------------
+
+def test_split_in_half_too_short_returns_empty():
+    assert split_in_half("court") == ("", "")
+
+
+def test_split_in_half_splits_on_newline_near_middle():
+    text = "A" * 300 + "\n" + "B" * 300
+    left, right = split_in_half(text)
+    assert left and right
+    assert left.startswith("A")
+    assert right.startswith("B")
+
+
+def test_split_in_half_halves_cover_all_content():
+    text = "\n".join(f"ligne {i} " + "x" * 20 for i in range(40))
+    left, right = split_in_half(text)
+    assert left and right
+    # Le découpage ne perd rien : la concaténation contient début et fin.
+    assert "ligne 0" in left
+    assert "ligne 39" in right

brain/tests/test_embedding_adapters.py

@@ -0,0 +1,102 @@
+"""Tests des adapters d'embeddings (Mistral cloud + Ollama local)."""
+from __future__ import annotations
+
+import json
+
+import httpx
+import pytest
+import respx
+
+from app.application.embeddings import EmbeddingError
+from app.core.config import Settings
+from app.infrastructure.mistral_embedding_adapter import MistralEmbeddingProvider
+from app.infrastructure.ollama_embedding_adapter import OllamaEmbeddingProvider
+
+_MISTRAL = "https://api.mistral.ai/v1/embeddings"
+_OLLAMA = "http://ollama:11434/api/embed"
+
+
+def _settings(**kw) -> Settings:
+    base = dict(_env_file=None, llm_timeout_seconds=30, ollama_base_url="http://ollama:11434")
+    base.update(kw)
+    return Settings(**base)
+
+
+# --- Mistral -------------------------------------------------------------------
+
+def test_mistral_missing_key_raises_at_construction():
+    with pytest.raises(EmbeddingError):
+        MistralEmbeddingProvider(_settings(mistral_api_key=""))
+
+
+async def test_mistral_empty_texts_short_circuits():
+    svc = MistralEmbeddingProvider(_settings(mistral_api_key="k"))
+    assert await svc.embed([]) == []
+
+
+@respx.mock
+async def test_mistral_returns_vectors():
+    respx.post(_MISTRAL).mock(return_value=httpx.Response(200, json={
+        "data": [{"embedding": [0.1, 0.2]}, {"embedding": [0.3, 0.4]}]
+    }))
+    svc = MistralEmbeddingProvider(_settings(mistral_api_key="k", mistral_embedding_model="mistral-embed"))
+    vectors = await svc.embed(["texte un", "texte deux"])
+    assert vectors == [[0.1, 0.2], [0.3, 0.4]]
+
+
+@respx.mock
+async def test_mistral_http_error_raises():
+    respx.post(_MISTRAL).mock(return_value=httpx.Response(429, text="rate limit"))
+    svc = MistralEmbeddingProvider(_settings(mistral_api_key="k"))
+    with pytest.raises(EmbeddingError) as exc:
+        await svc.embed(["x"])
+    assert "429" in str(exc.value)
+
+
+@respx.mock
+async def test_mistral_size_mismatch_raises():
+    respx.post(_MISTRAL).mock(return_value=httpx.Response(200, json={"data": [{"embedding": [0.1]}]}))
+    svc = MistralEmbeddingProvider(_settings(mistral_api_key="k"))
+    with pytest.raises(EmbeddingError):
+        await svc.embed(["a", "b"])  # 2 demandés, 1 reçu
+
+
+# --- Ollama --------------------------------------------------------------------
+
+async def test_ollama_empty_texts_short_circuits():
+    svc = OllamaEmbeddingProvider(_settings(ollama_embedding_model="nomic-embed-text"))
+    assert await svc.embed([]) == []
+
+
+@respx.mock
+async def test_ollama_returns_vectors():
+    respx.post(_OLLAMA).mock(return_value=httpx.Response(200, json={"embeddings": [[0.1], [0.2]]}))
+    svc = OllamaEmbeddingProvider(_settings(ollama_embedding_model="mxbai-embed-large"))
+    assert await svc.embed(["a", "b"]) == [[0.1], [0.2]]
+
+
+@respx.mock
+async def test_ollama_applies_nomic_task_prefix():
+    route = respx.post(_OLLAMA).mock(return_value=httpx.Response(200, json={"embeddings": [[0.0]]}))
+    svc = OllamaEmbeddingProvider(_settings(ollama_embedding_model="nomic-embed-text"))
+    await svc.embed(["question ?"], kind="query")
+    sent = json.loads(route.calls.last.request.content)["input"]
+    assert sent == ["search_query: question ?"]
+
+
+@respx.mock
+async def test_ollama_no_prefix_for_non_nomic_model():
+    route = respx.post(_OLLAMA).mock(return_value=httpx.Response(200, json={"embeddings": [[0.0]]}))
+    svc = OllamaEmbeddingProvider(_settings(ollama_embedding_model="mxbai-embed-large"))
+    await svc.embed(["doc"], kind="document")
+    sent = json.loads(route.calls.last.request.content)["input"]
+    assert sent == ["doc"]
+
+
+@respx.mock
+async def test_ollama_http_error_mentions_pull_hint():
+    respx.post(_OLLAMA).mock(return_value=httpx.Response(404, text="model not found"))
+    svc = OllamaEmbeddingProvider(_settings(ollama_embedding_model="nomic-embed-text"))
+    with pytest.raises(EmbeddingError) as exc:
+        await svc.embed(["x"])
+    assert "ollama pull" in str(exc.value)

brain/tests/test_generate_page.py

@@ -0,0 +1,65 @@
+"""Tests du use case de génération de page (app.application.generate_page)."""
+from __future__ import annotations
+
+import pytest
+
+from app.application.generate_page import GeneratePageUseCase
+from app.domain.models import PageGenerationContext
+from app.domain.ports import LLMProviderError
+
+_CTX = PageGenerationContext(
+    lore_name="Eldoria",
+    folder_name="PNJ",
+    template_name="Personnage",
+    template_fields=["apparence", "histoire"],
+    page_title="Aragorn",
+    lore_description="un monde sombre",
+)
+
+
+def test_build_prompt_includes_context_and_fields():
+    p = GeneratePageUseCase._build_prompt(_CTX, "fr")
+    assert "Eldoria" in p
+    assert "Aragorn" in p
+    assert '"apparence"' in p
+    assert "un monde sombre" in p
+
+
+def test_build_prompt_omits_lore_description_when_absent():
+    ctx = PageGenerationContext("L", "F", "T", ["a"], "Titre", None)
+    assert "Description de l'univers" not in GeneratePageUseCase._build_prompt(ctx)
+
+
+def test_parse_values_keeps_only_expected_fields():
+    out = GeneratePageUseCase._parse_values(
+        '{"apparence":"grand","histoire":"longue","extra":"ignoré"}',
+        ["apparence", "histoire"])
+    assert out == {"apparence": "grand", "histoire": "longue"}
+
+
+def test_parse_values_missing_field_becomes_empty_string():
+    out = GeneratePageUseCase._parse_values('{"apparence":"grand"}', ["apparence", "histoire"])
+    assert out == {"apparence": "grand", "histoire": ""}
+
+
+def test_parse_values_casts_to_str_and_strips():
+    out = GeneratePageUseCase._parse_values('{"n": 42, "s": "  x  "}', ["n", "s"])
+    assert out == {"n": "42", "s": "x"}
+
+
+def test_parse_values_bad_json_raises():
+    with pytest.raises(LLMProviderError):
+        GeneratePageUseCase._parse_values("pas du json", ["a"])
+
+
+def test_parse_values_non_object_raises():
+    with pytest.raises(LLMProviderError):
+        GeneratePageUseCase._parse_values("[1, 2]", ["a"])
+
+
+async def test_execute_returns_filtered_result():
+    class FakeLLM:
+        async def generate(self, prompt, *, output_format=None, temperature=None):
+            return '{"apparence":"grand","histoire":"épique","parasite":"x"}'
+    result = await GeneratePageUseCase(FakeLLM()).execute(_CTX)
+    assert result.values == {"apparence": "grand", "histoire": "épique"}

brain/tests/test_import_parsing.py

@@ -0,0 +1,95 @@
+"""Tests des parseurs robustes des use cases d'import (méthodes statiques).
+
+_parse_payload (campagne), _parse_sections / _parse_anchors (règles) : transforment
+la réponse brute du LLM en structure exploitable + un drapeau « tronqué » qui
+déclenche le re-découpage.
+"""
+from __future__ import annotations
+
+from app.application.import_campaign import ImportCampaignUseCase
+from app.application.import_rules import ImportRulesUseCase
+
+_parse_payload = ImportCampaignUseCase._parse_payload
+_parse_sections = ImportRulesUseCase._parse_sections
+_parse_anchors = ImportRulesUseCase._parse_anchors
+
+
+# --- campagne : _parse_payload -------------------------------------------------
+
+def test_parse_payload_valid():
+    payload, truncated = _parse_payload('{"arcs":[{"name":"A"}],"npcs":[{"name":"N"}]}', index=0)
+    assert truncated is False
+    assert payload == {"arcs": [{"name": "A"}], "npcs": [{"name": "N"}]}
+
+
+def test_parse_payload_truncated_flags_recut():
+    payload, truncated = _parse_payload('{"arcs":[{"name":"A"', index=0)
+    assert truncated is True
+    assert payload == {"arcs": [], "npcs": []}
+
+
+def test_parse_payload_prose_is_empty_not_truncated():
+    payload, truncated = _parse_payload('juste de la prose sans json', index=0)
+    assert truncated is False
+    assert payload == {"arcs": [], "npcs": []}
+
+
+def test_parse_payload_recovers_truncated_array():
+    raw = '{"arcs":[{"name":"A"},{"name":"B"},{"name":'
+    payload, truncated = _parse_payload(raw, index=0)
+    assert truncated is True
+    assert payload == {"arcs": [{"name": "A"}, {"name": "B"}], "npcs": []}
+
+
+def test_parse_payload_coerces_non_list_fields():
+    payload, _ = _parse_payload('{"arcs":"oops","npcs":null}', index=0)
+    assert payload == {"arcs": [], "npcs": []}
+
+
+# --- règles : _parse_sections --------------------------------------------------
+
+def test_parse_sections_valid_and_normalized():
+    sections, truncated = _parse_sections('{"sections":{"Combat":"texte"}}', index=0)
+    assert truncated is False
+    assert sections == {"Combat": "texte"}
+
+
+def test_parse_sections_truncated():
+    sections, truncated = _parse_sections('{"Combat":"texte non termin', index=0)
+    assert truncated is True
+    assert sections == {}
+
+
+def test_parse_sections_prose_is_empty():
+    sections, truncated = _parse_sections('pas de json ici', index=0)
+    assert truncated is False
+    assert sections == {}
+
+
+# --- règles (mode segmentation) : _parse_anchors -------------------------------
+
+def test_parse_anchors_locates_and_splits_text():
+    text = "Préambule.\nLE COMBAT commence ici, brutal.\nLA MAGIE ensuite, subtile."
+    raw = ('{"sections":[{"titre":"Combat","debut":"LE COMBAT commence"},'
+           '{"titre":"Magie","debut":"LA MAGIE ensuite"}]}')
+    sections, truncated = _parse_anchors(raw, text, index=0)
+    assert truncated is False
+    assert "Combat" in sections and "Magie" in sections
+    # La 1re section absorbe le préambule (avant la 1re ancre).
+    assert "Préambule." in sections["Combat"]
+    assert "LE COMBAT commence ici, brutal." in sections["Combat"]
+    assert "LA MAGIE ensuite, subtile." in sections["Magie"]
+
+
+def test_parse_anchors_unparseable_returns_empty():
+    sections, truncated = _parse_anchors("pas du json", "texte", index=0)
+    assert sections == {}
+    assert truncated is False
+
+
+def test_parse_anchors_anchor_not_found_is_dropped():
+    text = "Seulement ce paragraphe existe."
+    raw = '{"sections":[{"titre":"Fantôme","debut":"ancre absente du texte"}]}'
+    sections, _ = _parse_anchors(raw, text, index=0)
+    # Aucune ancre localisée → aucune section.
+    assert sections == {}

brain/tests/test_import_status.py

@@ -0,0 +1,30 @@
+"""Tests du canal de statut d'import (app.application.import_status)."""
+from __future__ import annotations
+
+import asyncio
+
+from app.application import import_status
+
+
+def test_notify_is_noop_without_queue():
+    # Hors import (aucune queue installée) : ne lève pas, ne fait rien.
+    import_status.notify_status("personne n'écoute")  # ne doit pas lever
+
+
+def test_notify_publishes_when_queue_installed():
+    queue: asyncio.Queue = asyncio.Queue()
+    token = import_status.set_status_queue(queue)
+    try:
+        import_status.notify_status("morceau re-découpé")
+        assert queue.get_nowait() == "morceau re-découpé"
+    finally:
+        import_status.reset_status_queue(token)
+
+
+def test_reset_restores_noop():
+    queue: asyncio.Queue = asyncio.Queue()
+    token = import_status.set_status_queue(queue)
+    import_status.reset_status_queue(token)
+    # Après reset : plus de queue active → no-op, la queue reste vide.
+    import_status.notify_status("ignoré")
+    assert queue.empty()

brain/tests/test_import_use_cases.py

@@ -0,0 +1,157 @@
+"""Tests des use cases d'import via FAKES (ports LLM + extracteur PDF).
+
+Exerce la chaîne map-reduce complète (extraction → chunking → MAP → REDUCE →
+streaming d'événements) SANS réseau ni vrai PDF. `chunk_text` est monkeypatché
+pour un découpage déterministe (le chunking est testé à part). `asyncio.sleep`
+est neutralisé pour que les backoffs de retry n'imposent aucune attente.
+"""
+from __future__ import annotations
+
+import pytest
+
+from app.application.import_campaign import ImportCampaignUseCase
+from app.application.import_rules import ImportRulesUseCase
+from app.domain.models import ExtractedDocument, ExtractedPage
+from app.domain.ports import LLMProviderError
+
+
+# --- fakes ---------------------------------------------------------------------
+
+class FakeExtractor:
+    def __init__(self, doc: ExtractedDocument) -> None:
+        self._doc = doc
+
+    def extract(self, pdf_bytes: bytes) -> ExtractedDocument:
+        return self._doc
+
+
+class ScriptedLLM:
+    """Rejoue une réponse par appel (la dernière est répétée si on dépasse)."""
+
+    def __init__(self, responses: list) -> None:
+        self._responses = list(responses)
+        self.calls = 0
+
+    async def generate(self, prompt: str, *, output_format=None, temperature=None) -> str:
+        r = self._responses[min(self.calls, len(self._responses) - 1)]
+        self.calls += 1
+        if isinstance(r, Exception):
+            raise r
+        return r
+
+
+class ContentLLM:
+    """Répond selon le CONTENU du prompt (chunk) : (sous-chaîne → réponse/exception)."""
+
+    def __init__(self, rules: list) -> None:
+        self._rules = rules
+
+    async def generate(self, prompt: str, *, output_format=None, temperature=None) -> str:
+        for sub, r in self._rules:
+            if sub in prompt:
+                if isinstance(r, Exception):
+                    raise r
+                return r
+        raise AssertionError(f"aucune règle ContentLLM ne matche : {prompt[:60]!r}")
+
+
+def _doc(text: str = "Texte du PDF.", *, ocr: bool = False) -> ExtractedDocument:
+    return ExtractedDocument(pages=[ExtractedPage(index=0, text=text, used_ocr=ocr)])
+
+
+@pytest.fixture
+def no_sleep(monkeypatch):
+    async def _noop(_d):
+        return None
+    monkeypatch.setattr("asyncio.sleep", _noop)
+
+
+@pytest.fixture
+def one_chunk(monkeypatch):
+    monkeypatch.setattr("app.application.import_rules.chunk_text", lambda *a, **k: ["chunk"])
+    monkeypatch.setattr("app.application.import_campaign.chunk_text", lambda *a, **k: ["chunk"])
+
+
+# --- import de règles ----------------------------------------------------------
+
+async def test_rules_execute_returns_merged_sections(one_chunk):
+    llm = ScriptedLLM(['{"Combat":"## Combat\\nrègles de combat"}'])
+    uc = ImportRulesUseCase(llm, FakeExtractor(_doc(ocr=True)))
+    result = await uc.execute(b"pdf")
+    assert result.sections == {"Combat": "## Combat\nrègles de combat"}
+    assert result.page_count == 1
+    assert result.ocr_page_count == 1
+
+
+async def test_rules_stream_emits_extracting_start_progress_done(one_chunk):
+    llm = ScriptedLLM(['{"Magie":"sorts"}'])
+    uc = ImportRulesUseCase(llm, FakeExtractor(_doc()))
+    events = [e async for e in uc.stream(b"pdf")]
+    types = [e["type"] for e in events]
+    assert types[0] == "extracting"
+    assert types[1] == "start"
+    assert "progress" in types
+    done = events[-1]
+    assert done["type"] == "done"
+    assert done["sections"] == {"Magie": "sorts"}
+
+
+async def test_rules_stream_skips_failed_chunk_but_continues(monkeypatch, no_sleep):
+    monkeypatch.setattr("app.application.import_rules.chunk_text",
+                        lambda *a, **k: ["AAA premier", "BBB second"])
+    llm = ContentLLM([
+        ("AAA premier", LLMProviderError("HTTP 503 saturé")),
+        ("BBB second", '{"Magie":"sorts"}'),
+    ])
+    uc = ImportRulesUseCase(llm, FakeExtractor(_doc()))
+    events = [e async for e in uc.stream(b"pdf")]
+    types = [e["type"] for e in events]
+    assert "chunk_failed" in types
+    done = events[-1]
+    assert done["type"] == "done"
+    assert done["sections"] == {"Magie": "sorts"}
+    assert done["skipped"] == 1
+
+
+async def test_rules_stream_all_chunks_fail_emits_error(one_chunk, no_sleep):
+    llm = ScriptedLLM([LLMProviderError("HTTP 500 panne")])
+    uc = ImportRulesUseCase(llm, FakeExtractor(_doc()))
+    events = [e async for e in uc.stream(b"pdf")]
+    assert events[-1]["type"] == "error"
+    assert "échoué" in events[-1]["message"]
+
+
+# --- import de campagne --------------------------------------------------------
+
+_TREE = ('{"arcs":[{"name":"Acte I","description":"intro",'
+         '"chapters":[{"name":"Ch1","scenes":[{"name":"Sc1"}]}]}],'
+         '"npcs":[{"name":"Gandalf","description":"magicien"}]}')
+
+
+async def test_campaign_execute_builds_tree_and_npcs(one_chunk):
+    uc = ImportCampaignUseCase(ScriptedLLM([_TREE]), FakeExtractor(_doc()))
+    result = await uc.execute(b"pdf")
+    assert result.counts() == (1, 1, 1)
+    assert result.arcs[0].name == "Acte I"
+    assert result.arcs[0].chapters[0].scenes[0].name == "Sc1"
+    assert [n.name for n in result.npcs] == ["Gandalf"]
+
+
+async def test_campaign_stream_emits_done_with_serialized_tree(one_chunk):
+    uc = ImportCampaignUseCase(ScriptedLLM([_TREE]), FakeExtractor(_doc()))
+    events = [e async for e in uc.stream(b"pdf")]
+    types = [e["type"] for e in events]
+    assert types[0] == "extracting"
+    assert types[1] == "start"
+    assert "progress" in types
+    done = events[-1]
+    assert done["type"] == "done"
+    assert done["arcs"][0]["name"] == "Acte I"
+    assert done["arcs"][0]["chapters"][0]["scenes"][0]["name"] == "Sc1"
+    assert done["npcs"] == [{"name": "Gandalf", "description": "magicien"}]
+
+
+async def test_campaign_stream_all_fail_emits_error(one_chunk, no_sleep):
+    uc = ImportCampaignUseCase(ScriptedLLM([LLMProviderError("502")]), FakeExtractor(_doc()))
+    events = [e async for e in uc.stream(b"pdf")]
+    assert events[-1]["type"] == "error"

brain/tests/test_language.py

@@ -0,0 +1,39 @@
+"""Tests de la normalisation de langue (app.core.language)."""
+from __future__ import annotations
+
+import pytest
+
+from app.core import language
+
+
+@pytest.mark.parametrize("raw, expected", [
+    ("fr", "fr"),
+    ("en", "en"),
+    ("EN", "en"),
+    ("en-US", "en"),
+    ("fr-FR,fr;q=0.9,en;q=0.8", "fr"),
+    ("en-GB,en;q=0.9", "en"),
+    ("de", "fr"),          # non supporté → défaut
+    ("", "fr"),
+    (None, "fr"),
+    ("  EN-gb ", "en"),    # casse + espaces tolérés
+])
+def test_normalize(raw, expected):
+    assert language.normalize(raw) == expected
+
+
+def test_language_name_known_and_fallback():
+    assert language.language_name("fr") == "français"
+    assert language.language_name("en") == "anglais"
+    # Code inconnu → nom de la langue par défaut.
+    assert language.language_name("xx") == "français"
+
+
+def test_instruction_mentions_target_language():
+    assert "anglais" in language.instruction("en")
+    assert "français" in language.instruction("fr")
+
+
+def test_get_user_language_uses_normalize():
+    assert language.get_user_language("en-US") == "en"
+    assert language.get_user_language(None) == "fr"

brain/tests/test_llm_json.py

@@ -0,0 +1,130 @@
+"""Tests de la lecture robuste de JSON depuis une réponse LLM (app.application.llm_json).
+
+Couvre l'extraction du premier objet équilibré (en ignorant les accolades dans
+les chaînes), la réparation d'un JSON tronqué, la détection « ça ressemble à du
+JSON coupé », et le strip des blocs de raisonnement <think>…</think>.
+"""
+from __future__ import annotations
+
+from app.application.llm_json import (
+    extract_json_object,
+    load_json_object,
+    looks_like_truncated_json,
+    repair_truncated_json,
+)
+
+
+# --- extract_json_object -------------------------------------------------------
+
+def test_extract_simple_object():
+    assert extract_json_object('{"a": 1}') == '{"a": 1}'
+
+
+def test_extract_ignores_surrounding_prose_and_fences():
+    raw = 'Voici le JSON :\n```json\n{"a": 1}\n```\nMerci.'
+    assert extract_json_object(raw) == '{"a": 1}'
+
+
+def test_extract_stops_at_first_balanced_object():
+    assert extract_json_object('{"a": 1} et puis {"b": 2}') == '{"a": 1}'
+
+
+def test_extract_keeps_nested_object_whole():
+    assert extract_json_object('{"a": {"b": 1}}') == '{"a": {"b": 1}}'
+
+
+def test_extract_ignores_braces_inside_strings():
+    raw = '{"a": "}{ pas du json "}'
+    assert extract_json_object(raw) == raw
+
+
+def test_extract_handles_escaped_quote_in_string():
+    raw = '{"a": "x\\"y"}'
+    assert extract_json_object(raw) == raw
+
+
+def test_extract_returns_none_when_unclosed():
+    assert extract_json_object('{"a": 1') is None
+
+
+def test_extract_returns_none_without_brace():
+    assert extract_json_object('aucune accolade ici') is None
+
+
+def test_extract_returns_none_on_empty():
+    assert extract_json_object('') is None
+
+
+# --- load_json_object ----------------------------------------------------------
+
+def test_load_valid_object_not_recovered():
+    obj, recovered = load_json_object('{"x": 42}')
+    assert obj == {"x": 42}
+    assert recovered is False
+
+
+def test_load_tolerates_raw_control_chars_in_strings():
+    # Retour à la ligne BRUT dans une chaîne : invalide en strict, accepté ici.
+    obj, recovered = load_json_object('{"a": "ligne1\nligne2"}')
+    assert obj == {"a": "ligne1\nligne2"}
+    assert recovered is False
+
+
+def test_load_strips_reasoning_block_before_parsing():
+    raw = '<think>je réfléchis { ] [ }</think>{"ok": true}'
+    obj, recovered = load_json_object(raw)
+    assert obj == {"ok": True}
+    assert recovered is False
+
+
+def test_load_repairs_truncated_array_and_flags_recovered():
+    raw = '{"items": [{"a": 1}, {"b": 2}, {"c":'
+    obj, recovered = load_json_object(raw)
+    assert obj == {"items": [{"a": 1}, {"b": 2}]}
+    assert recovered is True
+
+
+def test_load_returns_none_on_garbage():
+    obj, recovered = load_json_object('juste de la prose sans json')
+    assert obj is None
+    assert recovered is False
+
+
+# --- looks_like_truncated_json -------------------------------------------------
+
+def test_truncated_detection_no_brace_is_false():
+    assert looks_like_truncated_json('rien') is False
+
+
+def test_truncated_detection_short_object_start_unbalanced_is_true():
+    # Démarre par '{' et déséquilibré → coupé net, même très court.
+    assert looks_like_truncated_json('{"') is True
+
+
+def test_truncated_detection_balanced_object_is_false():
+    assert looks_like_truncated_json('{"a": 1}') is False
+
+
+def test_truncated_detection_short_prose_with_braces_is_false():
+    assert looks_like_truncated_json('texte { incomplet') is False
+
+
+def test_truncated_detection_long_prose_unbalanced_is_true():
+    raw = 'prose ' * 30 + '{ structure ouverte mais jamais refermée'
+    assert len(raw) >= 100
+    assert looks_like_truncated_json(raw) is True
+
+
+# --- repair_truncated_json -----------------------------------------------------
+
+def test_repair_closes_open_containers_after_last_complete_element():
+    repaired = repair_truncated_json('{"items": [{"a": 1}, {"b": 2}, {"c":')
+    assert repaired == '{"items": [{"a": 1}, {"b": 2}]}'
+
+
+def test_repair_returns_none_when_nothing_complete():
+    assert repair_truncated_json('{"a": "jamais fermé') is None
+
+
+def test_repair_returns_none_without_brace():
+    assert repair_truncated_json('pas de json') is None

brain/tests/test_llm_retry.py

@@ -0,0 +1,112 @@
+"""Tests du retry des appels LLM one-shot (app.application.llm_retry).
+
+`asyncio.sleep` est neutralisé (et enregistré) pour que les backoffs n'imposent
+aucune attente réelle tout en vérifiant les durées choisies.
+"""
+from __future__ import annotations
+
+import pytest
+
+from app.application import llm_retry
+from app.application.llm_retry import (
+    _is_daily_quota,
+    _is_rate_limit,
+    _suggested_retry_after,
+    generate_with_retry,
+)
+from app.domain.ports import LLMGenerationTimeout, LLMProviderError
+
+
+class FakeLLM:
+    """LLM factice : rejoue une liste de comportements (exception ou texte)."""
+
+    def __init__(self, behaviors: list) -> None:
+        self._behaviors = list(behaviors)
+        self.calls = 0
+
+    async def generate(self, prompt: str, *, output_format=None, temperature=None) -> str:
+        b = self._behaviors[self.calls]
+        self.calls += 1
+        if isinstance(b, Exception):
+            raise b
+        return b
+
+
+@pytest.fixture
+def slept(monkeypatch):
+    """Neutralise asyncio.sleep et enregistre les durées demandées."""
+    recorded: list[float] = []
+
+    async def fake_sleep(d):
+        recorded.append(d)
+
+    monkeypatch.setattr("asyncio.sleep", fake_sleep)
+    return recorded
+
+
+# --- helpers de classification -------------------------------------------------
+
+def test_is_rate_limit():
+    assert _is_rate_limit(LLMProviderError("HTTP 429 Too Many Requests"))
+    assert _is_rate_limit(LLMProviderError("rate limit reached"))
+    assert not _is_rate_limit(LLMProviderError("HTTP 500"))
+
+
+def test_is_daily_quota():
+    assert _is_daily_quota(LLMProviderError("free-models-per-day limit"))
+    assert _is_daily_quota(LLMProviderError("quota per day exceeded"))
+    assert not _is_daily_quota(LLMProviderError("429 per-minute"))
+
+
+def test_suggested_retry_after():
+    assert _suggested_retry_after(LLMProviderError('{"retry_after_seconds": 8}')) == 8.0
+    assert _suggested_retry_after(LLMProviderError('Retry-After: 12')) == 12.0
+    assert _suggested_retry_after(LLMProviderError("pas de hint")) is None
+
+
+# --- generate_with_retry -------------------------------------------------------
+
+async def test_returns_on_first_success(slept):
+    llm = FakeLLM(["réponse"])
+    assert await generate_with_retry(llm, "p") == "réponse"
+    assert llm.calls == 1
+    assert slept == []
+
+
+async def test_retries_transient_error_then_succeeds(slept):
+    llm = FakeLLM([LLMProviderError("HTTP 503"), "ok"])
+    assert await generate_with_retry(llm, "p") == "ok"
+    assert llm.calls == 2
+    assert slept == [3.0]  # _BASE_DELAY_SECONDS
+
+
+async def test_timeout_raises_immediately_without_retry(slept):
+    llm = FakeLLM([LLMGenerationTimeout("trop lent")])
+    with pytest.raises(LLMGenerationTimeout):
+        await generate_with_retry(llm, "p")
+    assert llm.calls == 1
+    assert slept == []
+
+
+async def test_daily_quota_aborts_immediately(slept):
+    llm = FakeLLM([LLMProviderError("free-models-per-day exceeded")])
+    with pytest.raises(LLMProviderError):
+        await generate_with_retry(llm, "p")
+    assert llm.calls == 1
+    assert slept == []
+
+
+async def test_exhausts_attempts_then_raises_last(slept):
+    llm = FakeLLM([LLMProviderError("503 a"), LLMProviderError("503 b"), LLMProviderError("503 c")])
+    with pytest.raises(LLMProviderError, match="503 c"):
+        await generate_with_retry(llm, "p")
+    assert llm.calls == 3
+    # 2 attentes entre 3 tentatives (backoff exponentiel 3s puis 6s).
+    assert slept == [3.0, 6.0]
+
+
+async def test_rate_limit_respects_suggested_retry_after(slept):
+    llm = FakeLLM([LLMProviderError('429 {"retry_after_seconds": 8}'), "ok"])
+    assert await generate_with_retry(llm, "p") == "ok"
+    # min(8 + 2, 60) = 10
+    assert slept == [10.0]

brain/tests/test_models.py

@@ -0,0 +1,52 @@
+"""Tests de la logique portée par les modèles de domaine (app.domain.models)."""
+from __future__ import annotations
+
+from app.domain.models import (
+    ArcProposal,
+    CampaignImportResult,
+    ChapterProposal,
+    ExtractedDocument,
+    ExtractedPage,
+    RulesImportResult,
+    SceneProposal,
+)
+
+
+def test_extracted_document_properties():
+    doc = ExtractedDocument(pages=[
+        ExtractedPage(index=0, text="page un", used_ocr=False),
+        ExtractedPage(index=1, text="page deux", used_ocr=True),
+        ExtractedPage(index=2, text="   ", used_ocr=False),  # vide → exclue de full_text
+    ])
+    assert doc.page_count == 3
+    assert doc.ocr_page_count == 1
+    assert doc.full_text == "page un\n\npage deux"
+
+
+def test_rules_import_result_to_markdown():
+    result = RulesImportResult(
+        sections={"Combat": "règles de combat", "Magie": "règles de magie"},
+        page_count=10, ocr_page_count=0,
+    )
+    md = result.to_markdown()
+    assert "## Combat\n\nrègles de combat" in md
+    assert "## Magie\n\nrègles de magie" in md
+    assert md.endswith("\n")
+
+
+def test_campaign_import_result_counts():
+    arcs = [
+        ArcProposal("A1", "", chapters=[
+            ChapterProposal("C1", "", scenes=[SceneProposal("S1", ""), SceneProposal("S2", "")]),
+            ChapterProposal("C2", "", scenes=[SceneProposal("S3", "")]),
+        ]),
+        ArcProposal("A2", "", chapters=[]),
+    ]
+    result = CampaignImportResult(arcs=arcs, page_count=1, ocr_page_count=0)
+    assert result.counts() == (2, 2, 3)
+
+
+def test_arc_proposal_defaults():
+    arc = ArcProposal("Acte", "synopsis")
+    assert arc.arc_type == "LINEAR"
+    assert arc.chapters == []

brain/tests/test_ollama_adapter.py

@@ -0,0 +1,95 @@
+"""Tests de caractérisation de l'adapter Ollama (protocole propre : /api/generate
+one-shot + /api/chat NDJSON streamé)."""
+from __future__ import annotations
+
+import json
+
+import httpx
+import pytest
+import respx
+
+from app.core.config import Settings
+from app.domain.models import ChatMessage
+from app.domain.ports import LLMGenerationTimeout, LLMProviderError
+from app.infrastructure.ollama_adapter import OllamaLLMProvider
+
+_GEN = "http://ollama:11434/api/generate"
+_CHAT = "http://ollama:11434/api/chat"
+
+
+def _svc() -> OllamaLLMProvider:
+    s = Settings(_env_file=None, ollama_base_url="http://ollama:11434",
+                 llm_model="gemma", llm_timeout_seconds=30, llm_num_ctx=8192)
+    return OllamaLLMProvider(s)
+
+
+@respx.mock
+async def test_generate_returns_response_field():
+    respx.post(_GEN).mock(return_value=httpx.Response(200, json={"response": "texte", "done_reason": "stop"}))
+    assert await _svc().generate("prompt") == "texte"
+
+
+@respx.mock
+async def test_generate_payload_always_sends_num_ctx_and_omits_temperature():
+    route = respx.post(_GEN).mock(return_value=httpx.Response(200, json={"response": "x"}))
+    await _svc().generate("p")
+    body = json.loads(route.calls.last.request.content)
+    assert body["model"] == "gemma"
+    assert body["stream"] is False
+    assert body["options"] == {"num_ctx": 8192}
+    assert "format" not in body
+
+
+@respx.mock
+async def test_generate_payload_includes_temperature_and_format_when_given():
+    route = respx.post(_GEN).mock(return_value=httpx.Response(200, json={"response": "x"}))
+    await _svc().generate("p", output_format="json", temperature=0.1)
+    body = json.loads(route.calls.last.request.content)
+    assert body["options"]["temperature"] == 0.1
+    assert body["format"] == "json"
+
+
+@respx.mock
+async def test_generate_http_error_surfaces_ollama_message():
+    respx.post(_GEN).mock(return_value=httpx.Response(404, json={"error": "model 'x' not found"}))
+    with pytest.raises(LLMProviderError) as exc:
+        await _svc().generate("p")
+    assert "not found" in str(exc.value)
+    assert "404" in str(exc.value)
+
+
+@respx.mock
+async def test_generate_read_timeout_is_generation_timeout():
+    respx.post(_GEN).mock(side_effect=httpx.ReadTimeout("trop lent"))
+    with pytest.raises(LLMGenerationTimeout):
+        await _svc().generate("p")
+
+
+@respx.mock
+async def test_generate_connect_timeout_is_provider_error():
+    respx.post(_GEN).mock(side_effect=httpx.ConnectTimeout("injoignable"))
+    with pytest.raises(LLMProviderError) as exc:
+        await _svc().generate("p")
+    assert not isinstance(exc.value, LLMGenerationTimeout)
+
+
+@respx.mock
+async def test_stream_chat_yields_tokens_until_done():
+    body = (
+        '{"message":{"content":"Bon"},"done":false}\n'
+        '{"message":{"content":"jour"},"done":false}\n'
+        '{"done":true}\n'
+    )
+    respx.post(_CHAT).mock(return_value=httpx.Response(200, text=body))
+    tokens = [t async for t in _svc().stream_chat([ChatMessage(role="user", content="hi")])]
+    assert tokens == ["Bon", "jour"]
+
+
+@respx.mock
+async def test_stream_chat_prepends_system_prompt():
+    route = respx.post(_CHAT).mock(return_value=httpx.Response(200, text='{"done":true}\n'))
+    _ = [t async for t in _svc().stream_chat(
+        [ChatMessage(role="user", content="Q")], system_prompt="SYS")]
+    body = json.loads(route.calls.last.request.content)
+    assert body["messages"][0] == {"role": "system", "content": "SYS"}
+    assert body["messages"][-1] == {"role": "user", "content": "Q"}

brain/tests/test_onemin_adapter.py

@@ -0,0 +1,110 @@
+"""Tests de l'adapter 1min.ai (API propriétaire : prompt unique aplati, SSE
+`event: content`/`data:{content}`)."""
+from __future__ import annotations
+
+import httpx
+import pytest
+import respx
+
+from app.core.config import Settings
+from app.domain.models import ChatMessage
+from app.domain.ports import LLMProviderError
+from app.infrastructure.onemin_adapter import OneMinAiLLMProvider
+
+_URL = "https://api.1min.ai/api/chat-with-ai?isStreaming=true"
+
+
+def _svc() -> OneMinAiLLMProvider:
+    s = Settings(_env_file=None, onemin_api_key="k", onemin_model="gpt-4o-mini",
+                 llm_timeout_seconds=30)
+    return OneMinAiLLMProvider(s)
+
+
+def _sse(*blocks: str) -> str:
+    return "".join(blocks)
+
+
+# --- streaming -----------------------------------------------------------------
+
+@respx.mock
+async def test_generate_collects_content_chunks():
+    body = _sse(
+        "event: content\ndata: {\"content\": \"Bon\"}\n\n",
+        "event: content\ndata: {\"content\": \"jour\"}\n\n",
+        "event: done\ndata: {}\n\n",
+    )
+    respx.post(_URL).mock(return_value=httpx.Response(200, text=body))
+    assert await _svc().generate("salut") == "Bonjour"
+
+
+@respx.mock
+async def test_generate_sends_api_key_header_and_prompt_payload():
+    route = respx.post(_URL).mock(return_value=httpx.Response(
+        200, text="event: done\ndata: {}\n\n"))
+    await _svc().generate("ma question")
+    req = route.calls.last.request
+    assert req.headers["API-KEY"] == "k"
+    import json
+    body = json.loads(req.content)
+    assert body["model"] == "gpt-4o-mini"
+    assert body["promptObject"]["prompt"] == "ma question"
+
+
+@respx.mock
+async def test_error_event_raises_provider_error():
+    body = "event: error\ndata: {\"message\": \"quota dépassé\"}\n\n"
+    respx.post(_URL).mock(return_value=httpx.Response(200, text=body))
+    with pytest.raises(LLMProviderError) as exc:
+        await _svc().generate("p")
+    assert "quota dépassé" in str(exc.value)
+
+
+@respx.mock
+async def test_http_error_is_translated():
+    respx.post(_URL).mock(return_value=httpx.Response(502, text="bad gateway"))
+    with pytest.raises(LLMProviderError) as exc:
+        await _svc().generate("p")
+    assert "1min.ai" in str(exc.value)
+
+
+@respx.mock
+async def test_stream_chat_flattens_and_streams():
+    route = respx.post(_URL).mock(return_value=httpx.Response(
+        200, text="event: content\ndata: {\"content\": \"R\"}\n\nevent: done\ndata: {}\n\n"))
+    tokens = [t async for t in _svc().stream_chat(
+        [ChatMessage(role="user", content="Q")], system_prompt="SYS")]
+    assert tokens == ["R"]
+    import json
+    prompt = json.loads(route.calls.last.request.content)["promptObject"]["prompt"]
+    assert "[SYSTEM]" in prompt and "SYS" in prompt
+    assert "[USER]" in prompt and "Q" in prompt
+
+
+# --- helpers purs --------------------------------------------------------------
+
+def test_flatten_messages_structure():
+    out = OneMinAiLLMProvider._flatten_messages(
+        [ChatMessage(role="user", content="Q1"), ChatMessage(role="assistant", content="R1")],
+        "instructions système",
+    )
+    assert "[SYSTEM]\ninstructions système" in out
+    assert "[USER]\nQ1" in out
+    assert "[ASSISTANT]\nR1" in out
+    assert out.rstrip().endswith("[ASSISTANT]")
+
+
+def test_extract_content_chunk_json_and_fallback():
+    assert OneMinAiLLMProvider._extract_content_chunk('{"content": "x"}') == "x"
+    assert OneMinAiLLMProvider._extract_content_chunk('{"token": "y"}') == "y"
+    # Non-JSON : filet de sécurité, on renvoie le brut.
+    assert OneMinAiLLMProvider._extract_content_chunk("texte brut") == "texte brut"
+
+
+def test_extract_result_reads_nested_result_object():
+    payload = {"aiRecord": {"aiRecordDetail": {"resultObject": ["partie 1", "partie 2"]}}}
+    assert OneMinAiLLMProvider._extract_result(payload) == "partie 1partie 2"
+
+
+def test_extract_result_raises_on_unexpected_schema():
+    with pytest.raises(LLMProviderError):
+        OneMinAiLLMProvider._extract_result({"unexpected": True})

brain/tests/test_openai_compatible_adapters.py

@@ -0,0 +1,204 @@
+"""Tests de caractérisation des adapters LLM « OpenAI-compatible »
+(OpenRouter, Gemini, Mistral).
+
+But : VERROUILLER le comportement observable AVANT d'extraire une classe de base
+commune (les trois adapters partageaient ~80 % de code). On couvre via respx
+(mock du transport httpx) : collecte du stream, payload envoyé, en-têtes, parsing
+SSE, et traduction des erreurs HTTP — sans aucun appel réseau réel.
+
+Ces tests doivent rester verts à l'identique après le refactor.
+"""
+from __future__ import annotations
+
+import json
+
+import httpx
+import pytest
+import respx
+
+from app.core.config import Settings
+from app.domain.models import ChatMessage
+from app.domain.ports import LLMProviderError
+from app.infrastructure.gemini_adapter import GeminiLLMProvider
+from app.infrastructure.mistral_adapter import MistralLLMProvider
+from app.infrastructure.openrouter_adapter import OpenRouterLLMProvider
+
+
+def _settings(**kw) -> Settings:
+    return Settings(_env_file=None, llm_timeout_seconds=30, **kw)
+
+
+def _sse(*contents: str) -> str:
+    """Construit un corps SSE OpenAI : une trame `data: {choices:[{delta:{content}}]}`
+    par fragment, terminé par `data: [DONE]`."""
+    lines: list[str] = []
+    for c in contents:
+        lines.append("data: " + json.dumps({"choices": [{"delta": {"content": c}}]}))
+        lines.append("")
+    lines += ["data: [DONE]", ""]
+    return "\n".join(lines)
+
+
+# (id, classe, url, kwargs settings (clé+modèle), supporte response_format=json_object)
+CASES = [
+    pytest.param(
+        OpenRouterLLMProvider,
+        "https://openrouter.ai/api/v1/chat/completions",
+        dict(openrouter_api_key="k", openrouter_model="m"),
+        False,
+        "OpenRouter",
+        id="openrouter",
+    ),
+    pytest.param(
+        GeminiLLMProvider,
+        "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions",
+        dict(gemini_api_key="k", gemini_model="m"),
+        True,
+        "Gemini",
+        id="gemini",
+    ),
+    pytest.param(
+        MistralLLMProvider,
+        "https://api.mistral.ai/v1/chat/completions",
+        dict(mistral_api_key="k", mistral_model="m"),
+        True,
+        "Mistral",
+        id="mistral",
+    ),
+]
+
+
+@pytest.mark.parametrize("cls, url, skw, supports_json, label", CASES)
+@respx.mock
+async def test_generate_collects_full_stream(cls, url, skw, supports_json, label):
+    respx.post(url).mock(return_value=httpx.Response(200, text=_sse("Bonjour", " le", " monde")))
+    svc = cls(_settings(**skw))
+    assert await svc.generate("salut") == "Bonjour le monde"
+
+
+@pytest.mark.parametrize("cls, url, skw, supports_json, label", CASES)
+@respx.mock
+async def test_stream_chat_yields_tokens(cls, url, skw, supports_json, label):
+    respx.post(url).mock(return_value=httpx.Response(200, text=_sse("A", "B", "C")))
+    svc = cls(_settings(**skw))
+    tokens = [t async for t in svc.stream_chat([ChatMessage(role="user", content="hi")])]
+    assert tokens == ["A", "B", "C"]
+
+
+@pytest.mark.parametrize("cls, url, skw, supports_json, label", CASES)
+@respx.mock
+async def test_payload_system_prompt_and_temperature(cls, url, skw, supports_json, label):
+    route = respx.post(url).mock(return_value=httpx.Response(200, text=_sse("x")))
+    svc = cls(_settings(**skw))
+    _ = [t async for t in svc.stream_chat(
+        [ChatMessage(role="user", content="Q")],
+        system_prompt="SYS",
+        temperature=0.5,
+    )]
+    body = json.loads(route.calls.last.request.content)
+    assert body["model"] == "m"
+    assert body["stream"] is True
+    assert body["messages"][0] == {"role": "system", "content": "SYS"}
+    assert body["messages"][-1] == {"role": "user", "content": "Q"}
+    assert body["temperature"] == 0.5
+
+
+@pytest.mark.parametrize("cls, url, skw, supports_json, label", CASES)
+@respx.mock
+async def test_payload_omits_temperature_when_none(cls, url, skw, supports_json, label):
+    route = respx.post(url).mock(return_value=httpx.Response(200, text=_sse("x")))
+    svc = cls(_settings(**skw))
+    await svc.generate("p")
+    body = json.loads(route.calls.last.request.content)
+    assert "temperature" not in body
+    # Sans system_prompt, generate envoie un unique message user.
+    assert body["messages"] == [{"role": "user", "content": "p"}]
+
+
+@pytest.mark.parametrize("cls, url, skw, supports_json, label", CASES)
+@respx.mock
+async def test_response_format_json_only_when_supported(cls, url, skw, supports_json, label):
+    route = respx.post(url).mock(return_value=httpx.Response(200, text=_sse("{}")))
+    svc = cls(_settings(**skw))
+    await svc.generate("p", output_format="json")
+    body = json.loads(route.calls.last.request.content)
+    if supports_json:
+        assert body["response_format"] == {"type": "json_object"}
+    else:
+        assert "response_format" not in body
+
+
+@pytest.mark.parametrize("cls, url, skw, supports_json, label", CASES)
+@respx.mock
+async def test_authorization_header_bearer(cls, url, skw, supports_json, label):
+    route = respx.post(url).mock(return_value=httpx.Response(200, text=_sse("x")))
+    svc = cls(_settings(**skw))
+    await svc.generate("p")
+    assert route.calls.last.request.headers["Authorization"] == "Bearer k"
+
+
+@respx.mock
+async def test_openrouter_attribution_headers():
+    route = respx.post("https://openrouter.ai/api/v1/chat/completions").mock(
+        return_value=httpx.Response(200, text=_sse("x")))
+    svc = OpenRouterLLMProvider(_settings(openrouter_api_key="k", openrouter_model="m"))
+    await svc.generate("p")
+    headers = route.calls.last.request.headers
+    assert headers["HTTP-Referer"] == "https://loremind.app"
+    assert headers["X-Title"] == "LoreMind"
+
+
+@pytest.mark.parametrize("cls, url, skw, supports_json, label", CASES)
+@respx.mock
+async def test_http_error_translated_to_provider_error(cls, url, skw, supports_json, label):
+    respx.post(url).mock(return_value=httpx.Response(429, text="quota exceeded"))
+    svc = cls(_settings(**skw))
+    with pytest.raises(LLMProviderError) as exc:
+        await svc.generate("p")
+    msg = str(exc.value)
+    assert label in msg
+    assert "429" in msg
+    assert "quota exceeded" in msg
+
+
+@respx.mock
+async def test_gemini_rejected_key_gives_actionable_message():
+    respx.post(
+        "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions"
+    ).mock(return_value=httpx.Response(403, text="API key not valid"))
+    svc = GeminiLLMProvider(_settings(gemini_api_key="k", gemini_model="m"))
+    with pytest.raises(LLMProviderError) as exc:
+        await svc.generate("p")
+    assert "refusée par Google" in str(exc.value)
+
+
+@pytest.mark.parametrize("cls, url, skw, supports_json, label", CASES)
+@respx.mock
+async def test_sse_skips_keepalive_and_malformed_lines(cls, url, skw, supports_json, label):
+    body = "\n".join([
+        ": OPENROUTER PROCESSING",  # commentaire keep-alive
+        "",
+        "data: not-json",            # JSON invalide -> ignoré
+        "",
+        "data: " + json.dumps({"choices": []}),  # pas de choix -> ignoré
+        "",
+        "data: " + json.dumps({"choices": [{"delta": {}}]}),  # delta sans content -> ignoré
+        "",
+        "data: " + json.dumps({"choices": [{"delta": {"content": "OK"}}]}),
+        "",
+        "data: [DONE]",
+        "",
+    ])
+    respx.post(url).mock(return_value=httpx.Response(200, text=body))
+    svc = cls(_settings(**skw))
+    assert await svc.generate("p") == "OK"
+
+
+@pytest.mark.parametrize("cls, skw", [
+    pytest.param(OpenRouterLLMProvider, dict(openrouter_api_key=""), id="openrouter"),
+    pytest.param(GeminiLLMProvider, dict(gemini_api_key=""), id="gemini"),
+    pytest.param(MistralLLMProvider, dict(mistral_api_key=""), id="mistral"),
+])
+def test_missing_api_key_raises_at_construction(cls, skw):
+    with pytest.raises(LLMProviderError):
+        cls(_settings(**skw))

brain/tests/test_query_rewrite.py

@@ -0,0 +1,54 @@
+"""Tests de la réécriture de question autonome (app.application.query_rewrite)."""
+from __future__ import annotations
+
+from app.application.query_rewrite import standalone_question
+from app.domain.models import ChatMessage
+
+
+class FakeLLM:
+    def __init__(self, response: str | None = None, exc: Exception | None = None) -> None:
+        self.response = response
+        self.exc = exc
+        self.called = False
+
+    async def generate(self, prompt, *, temperature=None, output_format=None) -> str:
+        self.called = True
+        if self.exc:
+            raise self.exc
+        return self.response
+
+
+async def test_single_turn_returns_last_user_without_calling_llm():
+    llm = FakeLLM()
+    q = await standalone_question(llm, [ChatMessage(role="user", content="Qui est Strahd ?")])
+    assert q == "Qui est Strahd ?"
+    assert llm.called is False
+
+
+async def test_multi_turn_uses_llm_rewrite_and_strips_quotes():
+    llm = FakeLLM(response='"Quelles sont les faiblesses de Strahd ?"')
+    msgs = [
+        ChatMessage(role="user", content="Qui est Strahd ?"),
+        ChatMessage(role="assistant", content="Un vampire."),
+        ChatMessage(role="user", content="Et ses faiblesses ?"),
+    ]
+    assert await standalone_question(llm, msgs) == "Quelles sont les faiblesses de Strahd ?"
+    assert llm.called is True
+
+
+async def test_llm_failure_falls_back_to_last_user():
+    llm = FakeLLM(exc=RuntimeError("LLM HS"))
+    msgs = [ChatMessage(role="user", content="A"), ChatMessage(role="user", content="B")]
+    assert await standalone_question(llm, msgs) == "B"
+
+
+async def test_suspiciously_long_rewrite_falls_back():
+    llm = FakeLLM(response="x" * 500)
+    msgs = [ChatMessage(role="user", content="A"), ChatMessage(role="user", content="B")]
+    assert await standalone_question(llm, msgs) == "B"
+
+
+async def test_empty_messages_returns_empty_string():
+    llm = FakeLLM()
+    assert await standalone_question(llm, []) == ""
+    assert llm.called is False

brain/tests/test_rerank.py

@@ -0,0 +1,60 @@
+"""Tests du reranking LLM des passages RAG (app.application.rerank)."""
+from __future__ import annotations
+
+from app.application.rerank import pool_size, rerank
+
+
+class FakeLLM:
+    def __init__(self, response: str | None = None, exc: Exception | None = None) -> None:
+        self.response = response
+        self.exc = exc
+
+    async def generate(self, prompt, *, temperature=None, output_format=None) -> str:
+        if self.exc:
+            raise self.exc
+        return self.response
+
+
+def test_pool_size():
+    assert pool_size(8) == 24          # min(max(24, 8), 24)
+    assert pool_size(4) == 12          # 4 * 3
+    assert pool_size(10) == 24         # plafonné à POOL_MAX
+    assert pool_size(1) == 3
+
+
+async def test_rerank_skips_when_pool_not_larger_than_top_k():
+    passages = [{"text": "a"}, {"text": "b"}]
+    # len <= top_k → renvoyé tel quel, sans appel LLM.
+    assert await rerank(FakeLLM(exc=AssertionError("ne doit pas être appelé")),
+                        "q", passages, top_k=3) == passages
+
+
+async def test_rerank_reorders_by_llm_scores():
+    passages = [{"text": "a"}, {"text": "b"}, {"text": "c"}]
+    out = await rerank(FakeLLM(response='{"scores":[1, 9, 5]}'), "q", passages, top_k=2)
+    assert [p["text"] for p in out] == ["b", "c"]
+
+
+async def test_rerank_stable_on_score_ties():
+    passages = [{"text": "a"}, {"text": "b"}, {"text": "c"}]
+    # Notes égales → ordre cosinus d'origine préservé.
+    out = await rerank(FakeLLM(response='{"scores":[5, 5, 5]}'), "q", passages, top_k=2)
+    assert [p["text"] for p in out] == ["a", "b"]
+
+
+async def test_rerank_llm_failure_falls_back_to_cosine_order():
+    passages = [{"text": "a"}, {"text": "b"}, {"text": "c"}]
+    out = await rerank(FakeLLM(exc=RuntimeError("LLM HS")), "q", passages, top_k=2)
+    assert [p["text"] for p in out] == ["a", "b"]
+
+
+async def test_rerank_wrong_score_count_falls_back():
+    passages = [{"text": "a"}, {"text": "b"}, {"text": "c"}]
+    out = await rerank(FakeLLM(response='{"scores":[1, 2]}'), "q", passages, top_k=2)
+    assert [p["text"] for p in out] == ["a", "b"]
+
+
+async def test_rerank_non_numeric_scores_fall_back():
+    passages = [{"text": "a"}, {"text": "b"}, {"text": "c"}]
+    out = await rerank(FakeLLM(response='{"scores":["x","y","z"]}'), "q", passages, top_k=2)
+    assert [p["text"] for p in out] == ["a", "b"]

brain/tests/test_section_merger.py

@@ -0,0 +1,107 @@
+"""Tests des helpers de l'import de règles (app.application.import_rules) :
+_SectionMerger, _normalize_sections, _coerce_markdown, _find_anchor, _combine_sections.
+"""
+from __future__ import annotations
+
+from app.application.import_rules import (
+    _SectionMerger,
+    _coerce_markdown,
+    _combine_sections,
+    _find_anchor,
+    _normalize_sections,
+)
+
+
+# --- _SectionMerger ------------------------------------------------------------
+
+def test_section_merger_case_insensitive_and_joins():
+    m = _SectionMerger()
+    touched = m.add({"Combat": "règle A", "combat": "règle B"})
+    assert touched == ["Combat"]            # clé canonique = 1re vue
+    res = m.result()
+    assert list(res.keys()) == ["Combat"]
+    assert res["Combat"] == "règle A\n\nrègle B"
+
+
+def test_section_merger_skips_empty_title_or_content():
+    m = _SectionMerger()
+    touched = m.add({"": "x", "Titre": "   ", "Vrai": "contenu"})
+    assert touched == ["Vrai"]
+    assert m.result() == {"Vrai": "contenu"}
+
+
+def test_section_merger_accumulates_across_chunks():
+    m = _SectionMerger()
+    m.add({"Combat": "p1"})
+    m.add({"Combat": "p2", "Magie": "sorts"})
+    res = m.result()
+    assert res["Combat"] == "p1\n\np2"
+    assert res["Magie"] == "sorts"
+
+
+# --- _normalize_sections -------------------------------------------------------
+
+def test_normalize_unwraps_known_envelope():
+    assert _normalize_sections({"sections": {"Combat": "x"}}) == {"Combat": "x"}
+    assert _normalize_sections({"règles": {"A": "y"}}) == {"A": "y"}
+
+
+def test_normalize_title_content_schema():
+    assert _normalize_sections({"title": "Combat", "content": "texte"}) == {"Combat": "texte"}
+
+
+def test_normalize_strips_meta_keys():
+    assert _normalize_sections({"Combat": "x", "thought": "bla", "notes": "y"}) == {"Combat": "x"}
+
+
+def test_normalize_passthrough_plain_sections():
+    assert _normalize_sections({"A": "1", "B": "2"}) == {"A": "1", "B": "2"}
+
+
+# --- _coerce_markdown ----------------------------------------------------------
+
+def test_coerce_markdown_string_passthrough():
+    assert _coerce_markdown("texte") == "texte"
+
+
+def test_coerce_markdown_none_is_empty():
+    assert _coerce_markdown(None) == ""
+
+
+def test_coerce_markdown_list_joined():
+    assert _coerce_markdown(["a", "b"]) == "a\n\nb"
+
+
+def test_coerce_markdown_dict_flattened():
+    out = _coerce_markdown({"Sous-titre": "contenu"})
+    assert "Sous-titre" in out
+    assert "contenu" in out
+
+
+# --- _find_anchor --------------------------------------------------------------
+
+def test_find_anchor_exact():
+    text = "Chapitre 1. Le héros entre."
+    assert _find_anchor(text, "Le héros entre", 0) == text.index("Le héros entre")
+
+
+def test_find_anchor_whitespace_flexible():
+    text = "Le   héros\nentre dans la taverne."
+    # Espaces multiples / saut de ligne dans le texte source, anchor normalisé.
+    assert _find_anchor(text, "Le héros entre dans la taverne", 0) is not None
+
+
+def test_find_anchor_case_insensitive():
+    assert _find_anchor("LE DONJON s'ouvre", "le donjon", 0) is not None
+
+
+def test_find_anchor_not_found():
+    assert _find_anchor("texte quelconque", "introuvable xyz", 0) is None
+
+
+# --- _combine_sections ---------------------------------------------------------
+
+def test_combine_sections_case_insensitive_concat():
+    out = _combine_sections({"Combat": "p1"}, {"combat": "p2", "Magie": "sorts"})
+    assert out["Combat"] == "p1\n\np2"
+    assert out["Magie"] == "sorts"

brain/tests/test_settings_store.py

@@ -0,0 +1,57 @@
+"""Tests des overrides runtime persistés (app.core.settings_store).
+
+Le chemin du fichier est redirigé vers un tmp_path pour isoler chaque test.
+"""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import pytest
+
+from app.core import settings_store
+
+
+@pytest.fixture(autouse=True)
+def isolated_store(tmp_path, monkeypatch):
+    monkeypatch.setattr(settings_store, "_OVERRIDES_PATH", tmp_path / "settings.json")
+    return tmp_path / "settings.json"
+
+
+def test_load_missing_file_returns_empty():
+    assert settings_store.load_overrides() == {}
+
+
+def test_save_filters_to_allowlist_and_persists(isolated_store):
+    result = settings_store.save_overrides({
+        "llm_model": "gemma3:12b",
+        "internal_shared_secret": "HACK",   # hors allow-list → ignoré
+        "champ_inconnu": "x",               # hors allow-list → ignoré
+    })
+    assert result == {"llm_model": "gemma3:12b"}
+    on_disk = json.loads(Path(isolated_store).read_text(encoding="utf-8"))
+    assert on_disk == {"llm_model": "gemma3:12b"}
+
+
+def test_save_merges_with_existing():
+    settings_store.save_overrides({"llm_model": "a"})
+    merged = settings_store.save_overrides({"llm_provider": "ollama"})
+    assert merged == {"llm_model": "a", "llm_provider": "ollama"}
+
+
+def test_load_ignores_non_allowlisted_keys_on_disk(isolated_store):
+    Path(isolated_store).write_text(
+        json.dumps({"llm_model": "ok", "internal_shared_secret": "leak"}),
+        encoding="utf-8",
+    )
+    assert settings_store.load_overrides() == {"llm_model": "ok"}
+
+
+def test_load_corrupted_file_returns_empty(isolated_store):
+    Path(isolated_store).write_text("{ pas du json", encoding="utf-8")
+    assert settings_store.load_overrides() == {}
+
+
+def test_load_non_dict_json_returns_empty(isolated_store):
+    Path(isolated_store).write_text("[1, 2, 3]", encoding="utf-8")
+    assert settings_store.load_overrides() == {}

brain/tests/test_streaming.py

@@ -0,0 +1,48 @@
+"""Tests des heartbeats SSE (app.application.streaming.with_heartbeat)."""
+from __future__ import annotations
+
+import asyncio
+
+import pytest
+
+from app.application.streaming import with_heartbeat
+
+
+async def _collect(agen) -> list[tuple[str, object]]:
+    return [ev async for ev in agen]
+
+
+async def test_fast_coro_emits_only_result():
+    async def quick() -> int:
+        return 42
+    events = await _collect(with_heartbeat(quick(), interval=0.05))
+    assert events == [("result", 42)]
+
+
+async def test_slow_coro_emits_heartbeats_then_result():
+    async def slow() -> str:
+        await asyncio.sleep(0.06)
+        return "fini"
+    events = await _collect(with_heartbeat(slow(), interval=0.02))
+    assert ("heartbeat", None) in events
+    assert events[-1] == ("result", "fini")
+
+
+async def test_relays_status_messages_from_queue():
+    queue: asyncio.Queue = asyncio.Queue()
+
+    async def work() -> str:
+        await asyncio.sleep(0.05)
+        return "ok"
+
+    queue.put_nowait("fournisseur saturé, nouvel essai")
+    events = await _collect(with_heartbeat(work(), interval=0.02, status_queue=queue))
+    assert ("status", "fournisseur saturé, nouvel essai") in events
+    assert events[-1] == ("result", "ok")
+
+
+async def test_propagates_coro_exception():
+    async def boom() -> None:
+        raise ValueError("échec interne")
+    with pytest.raises(ValueError, match="échec interne"):
+        await _collect(with_heartbeat(boom(), interval=0.05))

brain/tests/test_tree_merger.py

@@ -0,0 +1,149 @@
+"""Tests du _TreeMerger de l'import de campagne (app.application.import_campaign).
+
+Cœur du REDUCE : fusion par nom (insensible à la casse) des sous-arbres
+arc→chapitre→scène→pièce produits morceau par morceau, + accumulation des PNJ.
+"""
+from __future__ import annotations
+
+from app.application.import_campaign import _TreeMerger
+
+
+def test_single_chunk_builds_full_tree():
+    m = _TreeMerger()
+    m.add([{
+        "name": "Acte I", "description": "intro",
+        "chapters": [{
+            "name": "Ch1", "description": "d",
+            "scenes": [{
+                "name": "Sc1", "description": "s",
+                "player_narration": "PN", "gm_notes": "GM",
+                "rooms": [{"name": "R1", "description": "rd", "enemies": "gob", "loot": "or"}],
+            }],
+        }],
+    }])
+    arcs = m.result()
+    assert len(arcs) == 1
+    arc = arcs[0]
+    assert arc.name == "Acte I"
+    assert arc.arc_type == "LINEAR"
+    sc = arc.chapters[0].scenes[0]
+    assert sc.player_narration == "PN"
+    assert sc.gm_notes == "GM"
+    room = sc.rooms[0]
+    assert (room.name, room.enemies, room.loot) == ("R1", "gob", "or")
+
+
+def test_case_insensitive_arc_and_chapter_merge():
+    m = _TreeMerger()
+    m.add([{"name": "Acte I", "chapters": [{"name": "Ch1", "scenes": []}]}])
+    m.add([{"name": "acte i", "chapters": [{"name": "ch1", "scenes": []},
+                                           {"name": "Ch2", "scenes": []}]}])
+    arcs = m.result()
+    assert len(arcs) == 1
+    assert {c.name for c in arcs[0].chapters} == {"Ch1", "Ch2"}
+
+
+def test_description_first_non_empty_wins():
+    m = _TreeMerger()
+    m.add([{"name": "A", "description": "", "chapters": []}])
+    m.add([{"name": "A", "description": "vraie", "chapters": []}])
+    m.add([{"name": "A", "description": "autre", "chapters": []}])
+    assert m.result()[0].description == "vraie"
+
+
+def test_hub_type_wins_if_any_chunk_signals_it():
+    m = _TreeMerger()
+    m.add([{"name": "A", "type": "LINEAR", "chapters": []}])
+    m.add([{"name": "A", "type": "HUB", "chapters": []}])
+    assert m.result()[0].arc_type == "HUB"
+
+
+def _scene(narr=None, gm=None):
+    s = {"name": "S"}
+    if narr is not None:
+        s["player_narration"] = narr
+    if gm is not None:
+        s["gm_notes"] = gm
+    return {"name": "A", "chapters": [{"name": "C", "scenes": [s]}]}
+
+
+def test_scene_narration_concatenated_across_chunks():
+    m = _TreeMerger()
+    m.add([_scene(narr="début")])
+    m.add([_scene(narr="suite")])
+    sc = m.result()[0].chapters[0].scenes[0]
+    assert sc.player_narration == "début\n\nsuite"
+
+
+def test_scene_field_dedups_exact_overlap():
+    m = _TreeMerger()
+    m.add([_scene(gm="texte identique")])
+    m.add([_scene(gm="texte identique")])
+    assert m.result()[0].chapters[0].scenes[0].gm_notes == "texte identique"
+
+
+def test_scene_field_takes_superset_version():
+    m = _TreeMerger()
+    m.add([_scene(gm="court")])
+    m.add([_scene(gm="court et bien plus long")])
+    assert m.result()[0].chapters[0].scenes[0].gm_notes == "court et bien plus long"
+
+
+def test_npcs_longest_description_wins():
+    m = _TreeMerger()
+    m.add_npcs([{"name": "Thorin", "description": "court"}])
+    m.add_npcs([{"name": "thorin", "description": "une description bien plus complète"}])
+    npcs = m.npcs()
+    assert len(npcs) == 1
+    assert npcs[0].name == "Thorin"
+    assert npcs[0].description == "une description bien plus complète"
+
+
+def test_counts():
+    m = _TreeMerger()
+    m.add([{"name": "A", "chapters": [
+        {"name": "C1", "scenes": [{"name": "S1"}, {"name": "S2"}]},
+        {"name": "C2", "scenes": []},
+    ]}])
+    assert m.counts() == (1, 2, 2)
+
+
+def test_blank_names_are_skipped():
+    m = _TreeMerger()
+    m.add([{"name": "", "chapters": []},
+           {"name": "   ", "chapters": []},
+           {"name": "OK", "chapters": [{"name": "", "scenes": []}]}])
+    arcs = m.result()
+    assert len(arcs) == 1
+    assert arcs[0].name == "OK"
+    assert arcs[0].chapters == []
+
+
+def test_merge_chapters_consolidation():
+    m = _TreeMerger()
+    m.add([{"name": "A", "chapters": [
+        {"name": "Intro", "scenes": [{"name": "S1"}]},
+        {"name": "Introduction", "scenes": [{"name": "S2"}]},
+    ]}])
+    assert m.merge_chapters("Intro", ["Introduction"]) is True
+    chapters = m.result()[0].chapters
+    assert len(chapters) == 1
+    assert {s.name for s in chapters[0].scenes} == {"S1", "S2"}
+
+
+def test_merge_chapters_unknown_target_returns_false():
+    m = _TreeMerger()
+    m.add([{"name": "A", "chapters": [{"name": "Intro", "scenes": []}]}])
+    assert m.merge_chapters("Inexistant", ["Intro"]) is False
+
+
+def test_merge_scenes_consolidation():
+    m = _TreeMerger()
+    m.add([{"name": "A", "chapters": [{"name": "C", "scenes": [
+        {"name": "Combat", "gm_notes": "x"},
+        {"name": "Le combat", "gm_notes": "y"},
+    ]}]}])
+    assert m.merge_scenes("C", "Combat", ["Le combat"]) is True
+    scenes = m.result()[0].chapters[0].scenes
+    assert len(scenes) == 1
+    assert scenes[0].name == "Combat"

brain/tests/test_vector_store.py

@@ -0,0 +1,119 @@
+"""Tests du stockage vectoriel fichier + recherche hybride (app.infrastructure.vector_store).
+
+Le répertoire de stockage est redirigé vers un tmp_path et le cache mémoire est
+vidé avant chaque test pour une isolation totale.
+"""
+from __future__ import annotations
+
+import pytest
+
+from app.infrastructure import vector_store
+
+
+@pytest.fixture(autouse=True)
+def isolated_store(tmp_path, monkeypatch):
+    monkeypatch.setattr(vector_store, "_STORE_DIR", tmp_path)
+    vector_store._CACHE.clear()
+    yield
+    vector_store._CACHE.clear()
+
+
+# --- cosinus -------------------------------------------------------------------
+
+def test_cosine_identical_is_one():
+    assert vector_store._cosine([1.0, 0.0], [2.0, 0.0]) == pytest.approx(1.0)
+
+
+def test_cosine_orthogonal_is_zero():
+    assert vector_store._cosine([1.0, 0.0], [0.0, 1.0]) == 0.0
+
+
+def test_cosine_mismatched_or_zero_is_zero():
+    assert vector_store._cosine([1.0], [1.0, 2.0]) == 0.0
+    assert vector_store._cosine([0.0, 0.0], [1.0, 1.0]) == 0.0
+    assert vector_store.cosine_similarity([], [1.0]) == 0.0  # alias public
+
+
+# --- mots significatifs --------------------------------------------------------
+
+def test_significant_words_filters_stopwords_and_short():
+    words = vector_store._significant_words("Le dragon DORT dans la caverne avec les gobelins")
+    assert "dragon" in words
+    assert "caverne" in words
+    assert "gobelins" in words
+    assert "les" not in words and "avec" not in words and "la" not in words
+
+
+# --- save / exists / delete ----------------------------------------------------
+
+def test_save_then_exists_and_delete():
+    vector_store.save("src1", ["chunk a"], [[1.0, 0.0]])
+    assert vector_store.exists("src1") is True
+    vector_store.delete("src1")
+    assert vector_store.exists("src1") is False
+
+
+def test_save_rejects_mismatched_lengths():
+    with pytest.raises(ValueError):
+        vector_store.save("s", ["a", "b"], [[1.0]])
+    with pytest.raises(ValueError):
+        vector_store.save("s", ["a"], [[1.0]], pages=[1, 2])
+
+
+def test_all_chunks_returns_text_and_page():
+    vector_store.save("s", ["t1", "t2"], [[1.0], [2.0]], pages=[3, 7])
+    chunks = vector_store.all_chunks("s")
+    assert chunks == [{"text": "t1", "page": 3}, {"text": "t2", "page": 7}]
+
+
+# --- recherche -----------------------------------------------------------------
+
+def test_search_ranks_by_cosine():
+    vector_store.save("s", ["proche", "loin"], [[1.0, 0.0], [0.0, 1.0]])
+    results = vector_store.search(["s"], [1.0, 0.0], top_k=2)
+    assert [r["text"] for r in results] == ["proche", "loin"]
+    assert results[0]["score"] > results[1]["score"]
+
+
+def test_search_respects_top_k():
+    vector_store.save("s", ["a", "b", "c"], [[1.0], [0.9], [0.8]])
+    assert len(vector_store.search(["s"], [1.0], top_k=2)) == 2
+
+
+def test_search_min_score_filters_out_weak_matches():
+    vector_store.save("s", ["proche", "orthogonal"], [[1.0, 0.0], [0.0, 1.0]])
+    results = vector_store.search(["s"], [1.0, 0.0], top_k=5, min_score=0.5)
+    assert [r["text"] for r in results] == ["proche"]
+
+
+def test_search_lexical_bonus_promotes_exact_term_match():
+    # Deux extraits de cosinus IDENTIQUE : le bonus lexical départage celui qui
+    # contient le mot exact de la question.
+    vector_store.save(
+        "s",
+        ["Strahd règne sur Barovia", "un texte neutre sans rapport"],
+        [[1.0, 0.0], [1.0, 0.0]],
+    )
+    results = vector_store.search(["s"], [1.0, 0.0], top_k=2, query_text="Strahd")
+    assert results[0]["text"] == "Strahd règne sur Barovia"
+    assert results[0]["score"] > results[1]["score"]
+
+
+def test_search_includes_source_id_and_page():
+    vector_store.save("livre", ["extrait"], [[1.0]], pages=[42])
+    [res] = vector_store.search(["livre"], [1.0], top_k=1)
+    assert res["source_id"] == "livre"
+    assert res["page"] == 42
+
+
+# --- résumés (analyse approfondie) ---------------------------------------------
+
+def test_summaries_roundtrip_keyed_by_batch_tokens():
+    vector_store.save_summaries("s", 1000, [{"summary": "résumé", "vector": [1.0]}])
+    assert vector_store.load_summaries("s", 1000) == [{"summary": "résumé", "vector": [1.0]}]
+    # Taille de lot différente → invalidé (le découpage ne correspondrait plus).
+    assert vector_store.load_summaries("s", 2000) is None
+
+
+def test_load_summaries_absent_returns_none():
+    assert vector_store.load_summaries("inconnu", 1000) is None