Let Gradio bind the port and graft the REST routes onto it

Our own uvicorn raced the runner for 7860; a bare ASGI app with no listener
left nothing serving. Gradio launches instead, and /asr, /translate and
/health are prepended to its router so its frontend catch-all cannot shadow
them. Health moves off / because Gradio owns / for the UI.

README.md

@@ -34,7 +34,7 @@ have to hold ~2GB of model weights in a 512MB process.
    - `HF_TOKEN` — a read token from the account that accepted the licence
    - `SERVICE_TOKEN` — any random string; the backend must send the same value
 
-The first request downloads weights and can take several minutes. `GET /`
+Gradio serves its status page at `/`. The first request downloads weights and can take several minutes. `GET /`
 answers immediately throughout and reports load state, so you can watch it come
 up without holding a request open.
 

app.py

@@ -88,7 +88,7 @@ def _auth(token):
         raise HTTPException(status_code=401, detail="bad service token")
 
 
-@app.get("/")
+@app.get("/health")
 def health():
     return {
         "status": "ok",
@@ -225,14 +225,14 @@ def translate(body: TranslateIn, x_service_token: str = Header(default=None)):
 
 
 # ------------------------------------------------------------------- LAUNCH ---
-# The Space runner serves the module-level `app` itself. Binding a port here as
-# well is what produced "address already in use" and killed the container: two
-# servers competing for 7860. So this module defines the ASGI app and never
-# listens.
+# A Gradio-SDK Space expects Gradio itself to bind the port. Two earlier shapes
+# both failed here: running our own uvicorn raced whatever the runner had
+# already put on 7860 ("address already in use"), and defining a bare ASGI app
+# with no listener left the container with nothing serving at all.
 #
-# The Blocks UI is grafted onto that same app at import time, which keeps /asr
-# and /translate at the root where the backend expects them, with a status page
-# at /ui.
+# So Gradio launches, and the REST routes are grafted onto the server it starts.
+# They are *prepended*, because Gradio registers catch-all routes for its own
+# frontend that would otherwise shadow /asr and /translate.
 def _ui_check():
     import json as _json
     return _json.dumps(health(), indent=2)
@@ -242,9 +242,10 @@ _UI_TEXT = (
     "## ScriptFlow inference service\n\n"
     "Hausa ASR + Hausa->English MT. The REST API is the real interface:\n\n"
     "- `POST /asr` — raw audio bytes as the body\n"
-    "- `POST /translate` — `{\"texts\": [...]}`\n\n"
-    "Both take an `X-Service-Token` header. The first call after a cold start "
-    "downloads ~1GB of weights and takes several minutes."
+    "- `POST /translate` — `{\"texts\": [...]}`\n"
+    "- `GET /health` — model load state\n\n"
+    "Both POSTs take an `X-Service-Token` header. The first call after a cold "
+    "start downloads ~1GB of weights and takes several minutes."
 )
 
 
@@ -252,13 +253,25 @@ def _build_ui():
     import gradio as gr
     with gr.Blocks(title="ScriptFlow Hausa inference") as demo:
         gr.Markdown(_UI_TEXT)
-        out = gr.Code(label="GET /", language="json")
+        out = gr.Code(label="GET /health", language="json")
         gr.Button("Check status").click(_ui_check, inputs=None, outputs=out)
     return demo
 
 
-try:
+if __name__ == "__main__":
     import gradio as gr
-    app = gr.mount_gradio_app(app, _build_ui(), path="/ui")
-except Exception as _e:                       # never let the UI break the API
-    print(f"gradio UI unavailable ({type(_e).__name__}: {_e}) — REST only", flush=True)
+
+    demo = _build_ui()
+    # prevent_thread_lock so this returns and the routes can be attached; the
+    # process is held open by block_thread() at the end instead.
+    demo.launch(
+        server_name="0.0.0.0",
+        server_port=int(os.environ.get("PORT", 7860)),
+        prevent_thread_lock=True,
+        show_error=True,
+    )
+    demo.app.router.routes[0:0] = app.router.routes
+    print("REST routes attached: " +
+          ", ".join(sorted(r.path for r in app.router.routes if hasattr(r, "methods"))),
+          flush=True)
+    demo.block_thread()