Runtime: faster System One server replies

MODEL_MANIFEST.json

@@ -569,7 +569,7 @@
    "d3_kernels.py": "8f7534ca1e1846ebf7c4c73058002f3be4dbf0d5d426f0815150ae0cbe7677bd",
    "modeling_d3.py": "f8f8db23ff58f1a990ffa24977d744a50e1b4446d02a64c22070967705b6dbef",
    "pipeline_d3.py": "4f9297c7c74e596e9b67df91a1e6823cfada48bf781336f141974ff6a2f1038e",
-   "d3_server.py": "934f5cd583176ed368e513484aa2aa854d378f948685114c9185b4c38d41e85d",
+   "d3_server.py": "766a20a69101e36ab0535a1ec683379db422380a83091f0eccd7e82edb6a02de",
    "d3_engine.py": "086238942762df45047a32cd9a8e2bedd9b63167eb40be260c776de567979f53",
    "requirements.txt": "de9defd71ae4e559a7f5bdcc935b23716f9788e8755f2ba6db1ab6862bcda12e",
    "d3_format.py": "3badfbc70561bb3f5b8ec16a11d2c455b96ddb32caace0a8a54ca102eae1f3fa"
@@ -605,7 +605,7 @@
   "d3_format.py": "3badfbc70561bb3f5b8ec16a11d2c455b96ddb32caace0a8a54ca102eae1f3fa",
   "d3_kernels.py": "8f7534ca1e1846ebf7c4c73058002f3be4dbf0d5d426f0815150ae0cbe7677bd",
   "d3_runtime.py": "3a38391f182b4f02988cf6f009eeb7dccd8d09a3617e686012e289b69cf5082a",
-  "d3_server.py": "934f5cd583176ed368e513484aa2aa854d378f948685114c9185b4c38d41e85d",
+  "d3_server.py": "766a20a69101e36ab0535a1ec683379db422380a83091f0eccd7e82edb6a02de",
   "decision_config.json": "6b80ca11bd6ba481df4b3c187db8a5d1983786563ca05e2edccb0bdac0d2fc4f",
   "merges.txt": "a9d356d7bdf1ef4949e3e748e95b8e10ad9d4e2e838eddc38a0a7b6b94d1db8d",
   "model-00001-of-00011.safetensors": "4ecb29a7d1e2d1d0ad97008e5405a11d510a8a712a0b6491928359176fd56029",
@@ -646,7 +646,7 @@
   "d3_format.py": 7008,
   "d3_kernels.py": 16129,
   "d3_runtime.py": 93977,
-  "d3_server.py": 17438,
+  "d3_server.py": 17705,
   "decision_config.json": 5503,
   "merges.txt": 3353259,
   "model-00001-of-00011.safetensors": 4997471976,

d3_server.py

@@ -172,6 +172,8 @@ class SystemOneServer(ThreadingHTTPServer):
     """One listening socket; a thread per connection; the model runs under the decider's lock."""
 
     daemon_threads = True
+    # Room for many clients connecting at once (the standard library's default backlog is 5).
+    request_queue_size = 128
 
     def __init__(self, address: tuple[str, int], decider: Decider, *, api_key: str | None = None):
         super().__init__(address, SystemOneHandler)
@@ -184,6 +186,8 @@ class SystemOneHandler(BaseHTTPRequestHandler):
     server_version = "d3-server"
     sys_version = ""
     timeout = IDLE_SECONDS
+    # Headers and body are separate writes; without TCP_NODELAY the body waits for the client's delayed ACK.
+    disable_nagle_algorithm = True
     # A request line the standard library cannot parse gets a plain-text error with a full status line.
     default_request_version = "HTTP/1.1"
     error_content_type = "text/plain; charset=utf-8"