Restore two-context interactive topology

Exact source: ttracx/thoxroute@49d114b14ba5f1a43fcb93d2f5a2b86a2e6e5e0c

Dockerfile

@@ -23,7 +23,7 @@ ENV HF_HOME=/home/mambauser/.cache/huggingface \
     THOX_FAST_MODEL_REPO=Qwen/Qwen2.5-0.5B-Instruct-GGUF \
     THOX_FAST_MODEL_REVISION=9217f5db79a29953eb74d5343926648285ec7e67 \
     THOX_FAST_MODEL_FILE=qwen2.5-0.5b-instruct-q4_k_m.gguf \
-    THOX_FAST_POOL_SIZE=1 \
+    THOX_FAST_POOL_SIZE=2 \
     THOX_FAST_QUEUE_TIMEOUT_S=6 \
     THOX_FAST_MAX_OUTPUT_TOKENS=16 \
     THOX_STREAM_HEARTBEAT_S=2 \

README.md

@@ -16,8 +16,10 @@ This Space exposes one OpenAI-compatible endpoint:
   a six-second queue ceiling;
 - `thox-rust-coder`: lazily loaded 25B/3B-active Rust-specialist target.
 
-`GET /healthz` is ready only after the interactive context is loaded and has
-completed a one-token startup warmup. Streaming sends a role chunk before model
+`GET /healthz` is ready only after both interactive contexts are loaded and have
+completed a one-token startup warmup. Each receives half the container CPU
+quota; this topology avoids the severe oversubscription regression observed
+with one eight-thread llama context. Streaming sends a role chunk before model
 evaluation so upstream time-to-first-byte watchdogs do not abandon healthy CPU
 work; two-second SSE comments keep the stream live during synchronous prompt
 evaluation. The interactive provider itself caps every request at 16 output

app.py

@@ -43,7 +43,7 @@ CODER_MODEL_FILE = os.environ.get(
     "THOX_CODER_MODEL_FILE", "Qwen3-Coder-REAP-25B-A3B-Rust-Q4_K_M.gguf"
 )
 N_CTX = int(os.environ.get("THOX_N_CTX", "4096"))
-FAST_POOL_SIZE = max(1, min(int(os.environ.get("THOX_FAST_POOL_SIZE", "1")), 2))
+FAST_POOL_SIZE = max(1, min(int(os.environ.get("THOX_FAST_POOL_SIZE", "2")), 2))
 FAST_QUEUE_TIMEOUT_S = max(
     0.1, min(float(os.environ.get("THOX_FAST_QUEUE_TIMEOUT_S", "6")), 10.0)
 )