perf: fix model reload thrash and cut time-to-first-token

Ollama reloaded the chat model mid-message because plan/summary/router
calls sent divergent runner options (no num_ctx/num_batch) and omitted
keep_alive — the main cause of perceived slowness.

- share one keep-alive httpx.AsyncClient for all Ollama calls
- unify runner options (runner_options) + keep_alive on every model
  call, embeddings included
- drop the blocking LLM routing fallback (pure lexical heuristic)
- run RAG recall + plan + code-model pick in parallel inside the SSE
  stream, after the start event
- RAG cosine scoring off the event loop; cache /api/tags 30s and
  nvidia-smi 5s; frontend polls 2s->5s, warm poll backoff, dedup
  config fetch

feat: working session menu in TopBar (switch/create/rename/delete)
feat: workspace file deletion (DELETE /api/files + UI trash buttons)
docs: recommended Ollama env vars (KEEP_ALIVE, MAX_LOADED_MODELS...)

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

backend/app/agent_config.py

@@ -179,15 +179,15 @@ def save_config(patch: dict, model: str | None = None) -> dict:
     return get_config(model)
 
 
-def ollama_options(cfg: dict) -> dict:
-    """Traduit la config en options de génération Ollama."""
-    opts = {
-        "temperature": cfg["temperature"],
-        "top_p": cfg["top_p"],
-        "top_k": cfg["top_k"],
-        "num_predict": cfg["max_tokens"],
-        "num_batch": cfg["num_batch"],
-    }
+def runner_options(cfg: dict) -> dict:
+    """Sous-ensemble d'options qui détermine l'identité du runner Ollama.
+
+    Ollama choisit son runner (processus de chargement du modèle) d'après
+    num_ctx / num_batch / num_gpu. TOUT appel au même modèle (plan, résumé,
+    agent…) doit envoyer ces mêmes valeurs, sinon Ollama recharge le modèle en
+    plein milieu d'un message — la cause principale des lenteurs observées.
+    """
+    opts: dict = {"num_batch": cfg["num_batch"]}
     # num_gpu n'est transmis que si l'utilisateur force explicitement un nombre
     # de couches (≥ 0). En -1 (défaut), on laisse Ollama auto-ajuster l'offload
     # GPU comme `ollama run` ; lui imposer une valeur peut le forcer sur le CPU.
@@ -200,5 +200,16 @@ def ollama_options(cfg: dict) -> dict:
     return opts
 
 
+def ollama_options(cfg: dict) -> dict:
+    """Traduit la config en options de génération Ollama."""
+    return {
+        **runner_options(cfg),
+        "temperature": cfg["temperature"],
+        "top_p": cfg["top_p"],
+        "top_k": cfg["top_k"],
+        "num_predict": cfg["max_tokens"],
+    }
+
+
 def enabled_tool_names(cfg: dict) -> list[str]:
     return [name for name, on in cfg["tools"].items() if on]