Add Légifrance API integration for live legal text injection

Implements a two-layer approach to ensure legal texts are accurate and current:

Option A — runtime injection (legifrance_client.py):
- LegifranceClient: OAuth2 via PISTE (oauth.piste.gouv.fr), fetches consolidated
  article text via api.piste.gouv.fr/dila/legifrance/lf-engine-app
- Redis cache with 24h TTL (in-memory fallback if no Redis)
- build_filter_prompt() injects live texts into the prompt with explicit priority
  over static descriptions — model uses official Légifrance text, not training memory
- Graceful fallback to static prompt if API is unavailable — filtering never blocks

Option B — weekly sync script (scripts/check_legal_refs.py):
- Reads legal_refs.yaml (11 tracked articles, priority critical/high/medium)
- Fetches current text from Légifrance, diffs against stored baseline
- --init: resolves LEGIARTI IDs on first run
- --check: report only; --update: saves new texts to YAML
- Exit code 2 when changes detected (CI/n8n-compatible)

Supporting changes:
- legal_refs.yaml: tracked articles with code LEGITEXT IDs, priorities, rationale
- requirements.txt: add pyyaml>=6.0.2 and requests>=2.32.0
- .env.example: document PISTE_CLIENT_ID, PISTE_CLIENT_SECRET, LEGIFRANCE_CACHE_TTL
- transparence.tsx: document the Légifrance injection for public transparency
- Fixed portal URL: piste.gouv.fr (not piste.api.gouv.fr — verified June 2026)

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

artifacts/flask-api/ai_agent.py

@@ -9,15 +9,33 @@ import json
 import os
 import logging
 from openai import OpenAI, BadRequestError
-from legal_framework import LEGAL_FILTER_PROMPT, SYNTHESIS_PROMPT, LAW_CHECK_PROMPT
+from legal_framework import LEGAL_FILTER_PROMPT, SYNTHESIS_PROMPT, LAW_CHECK_PROMPT, build_filter_prompt
+from legifrance_client import get_legifrance_client
 
 logger = logging.getLogger(__name__)
 
 MISTRAL_BASE_URL = "https://api.mistral.ai/v1"
+_LEGAL_REFS_PATH = os.path.join(os.path.dirname(__file__), "legal_refs.yaml")
 
 _client: OpenAI | None = None
 
 
+def _get_live_legal_texts() -> dict[str, str]:
+    """
+    Récupère les textes consolidés des articles critiques depuis Légifrance (PISTE).
+    Retourne {} si PISTE n'est pas configuré ou en cas d'erreur — sans bloquer le filtrage.
+    Les résultats sont mis en cache dans Redis (24h).
+    """
+    try:
+        lf = get_legifrance_client()
+        if not lf.available:
+            return {}
+        return lf.fetch_tracked_articles(_LEGAL_REFS_PATH)
+    except Exception:
+        logger.warning("Légifrance: impossible de récupérer les textes live — mode statique")
+        return {}
+
+
 def get_client() -> OpenAI:
     """
     Supporte deux modes (par ordre de priorité) :
@@ -58,12 +76,16 @@ def filter_idea(content: str) -> dict:
     try:
         client = get_client()
         filter_model = os.environ.get("FILTER_MODEL", os.environ.get("OPENAI_FILTER_MODEL", "mistral-small-latest"))
+        live_texts = _get_live_legal_texts()
+        system_prompt = build_filter_prompt(live_texts)
+        if live_texts:
+            logger.debug("Légifrance: %d article(s) injecté(s) dans le prompt", len(live_texts))
         response = client.chat.completions.create(
             model=filter_model,
             max_tokens=300,
             response_format={"type": "json_object"},
             messages=[
-                {"role": "system", "content": LEGAL_FILTER_PROMPT},
+                {"role": "system", "content": system_prompt},
                 {"role": "user", "content": f'Idée soumise : "{content}"'},
             ],
         )