feat(stt): mémoire long-terme sémantique via Qdrant (5b) (#10)

Serveur : longterm.py — collection Qdrant stt-memory (embeddings Qwen3 gpu-01, dim auto,
Cosine), recall top-k injecté au prompt, remember des tours user. Tout dégrade proprement
si Qdrant/embeddings injoignables (la mémoire court-terme tient). Env STT_MEMORY_LONGTERM,
STT_QDRANT_URL, STT_EMBED_URL, STT_MEMORY_TOPK.

Testé en process : dégradation OK (Qdrant down → mem=0, pas de crash, court-terme tient).
Qdrant réparé le 17/06 (5c). Recherche sémantique réelle à valider sur cluster.


Claude-Session: https://claude.ai/code/session_013FmcxGsyXZXogiAHQLjnZT

Co-authored-by: Claude <noreply@anthropic.com>

stt/server/stt_server/brain.py

@@ -12,8 +12,20 @@ import httpx
 from stt_server.config import settings
 
 
-async def ask(text: str, model: str | None = None, history: list[dict] | None = None) -> str:
-    messages = [{"role": "system", "content": settings.system_prompt}]
+async def ask(
+    text: str,
+    model: str | None = None,
+    history: list[dict] | None = None,
+    memories: list[str] | None = None,
+) -> str:
+    system = settings.system_prompt
+    if memories:
+        souvenirs = "\n".join(f"- {m}" for m in memories)
+        system += (
+            "\n\nÉléments de mémoire long-terme (peuvent aider, ignore si hors-sujet) :\n"
+            + souvenirs
+        )
+    messages = [{"role": "system", "content": system}]
     if history:
         messages += history
     messages.append({"role": "user", "content": text})