bench: defaults — parallel 8, tps-floor 15 scheduler-driven, stall timeout auto-scales with steps (bug-2222)

bench/package_courier/README.md

@@ -66,13 +66,22 @@ Offline graph-routing tests: `pytest test_agent_graph.py -q`.
 - **pooled** (default): the chat-templated system+tools prefix (~1.5k
   tokens) is pinned once as a PolyKV pool; every session attaches and
   reuses it (`cache_n >= prefix_len` on every turn). Worker count is
-  dynamic when `--tps-floor` is set: the scheduler grows the fleet
-  while per-session tps holds above the floor and shrinks it when the
-  engine saturates below it (live worker count and target are both
-  plotted in the report).
-- **naive** (`--no-polykv`): fixed worker fleet, no pool — every turn
-  re-prefills its full prompt. This is the baseline that shows what
-  prefix sharing is worth.
+  **dynamic by default** (`--tps-floor` defaults to **15**): the
+  scheduler grows the fleet while per-session tps holds above the
+  floor and shrinks it when the engine saturates below it (live worker
+  count and target are both plotted in the report). Pass
+  `--tps-floor 0` only when you deliberately want a fixed fleet.
+- **naive** (`--no-polykv`): fixed worker fleet (no scheduler — the
+  floor defaults to 0 here), no pool — every turn re-prefills its full
+  prompt. This is the baseline that shows what prefix sharing is worth.
+
+Defaults reflect that self-compensating settings are the point of the
+benchmark: engine `--parallel` defaults to **8**, the pooled worker
+fleet is floor-driven, and the work server's stall timeout auto-scales
+with chain length (`max(900 s, steps × 60 s)`; the timer resets on
+delivery, so long chains need the headroom — override with
+`--timeout`). Fixed worker counts are for the naive A/B leg, not for
+normal runs.
 
 ## Running
 

bench/package_courier/run.py

@@ -49,7 +49,7 @@ def main():
     ap.add_argument("--engine-url", help="attach to a running engine; skip boot")
     ap.add_argument("--engine-port", type=int, default=8231)
     ap.add_argument("--ctx", type=int, default=16384)
-    ap.add_argument("--parallel", type=int, default=4)
+    ap.add_argument("--parallel", type=int, default=8)
     ap.add_argument("--pools", type=int, default=8)
     ap.add_argument("--ctk", default="q8_0")
     ap.add_argument("--ctv", default="q8_0")
@@ -64,13 +64,17 @@ def main():
     ap.add_argument("--packages", type=int, default=100)
     ap.add_argument("--steps", type=int, default=10)
     ap.add_argument("--seed", type=int, default=1)
-    ap.add_argument("--timeout", type=float, default=900.0)
+    ap.add_argument("--timeout", type=float, default=None,
+                    help="work-server stall timeout in s (timer resets on "
+                         "delivery); default max(900, steps*60)")
     ap.add_argument("--work-port", type=int, default=8600)
     # agents
     ap.add_argument("--workers", type=int, default=4)
     ap.add_argument("--max-workers", type=int, default=16)
     ap.add_argument("--no-polykv", action="store_true")
-    ap.add_argument("--tps-floor", type=float, default=0.0)
+    ap.add_argument("--tps-floor", type=float, default=None,
+                    help="worker scheduler tps floor; default 15 (pooled). "
+                         "Naive mode has no scheduler: default 0 (fixed fleet)")
     ap.add_argument("--temperature", type=float, default=0.0)
     ap.add_argument("--max-tokens", type=int, default=256)
     # output
@@ -81,6 +85,10 @@ def main():
                     help="just regenerate report.html from --out's existing run "
                          "data (no engine/bench); same as: report_html.py <out>/run")
     cfg = ap.parse_args()
+    if cfg.tps_floor is None:
+        cfg.tps_floor = 0.0 if cfg.no_polykv else 15.0
+    if cfg.timeout is None:
+        cfg.timeout = max(900.0, cfg.steps * 60.0)
 
     if cfg.report_only:
         from report_html import build

bench/package_courier/work_server.py

@@ -260,9 +260,13 @@ def main():
     ap.add_argument("--packages", type=int, default=100)
     ap.add_argument("--steps", type=int, default=10)
     ap.add_argument("--seed", type=int, default=1)
-    ap.add_argument("--timeout", type=float, default=900.0)
+    ap.add_argument("--timeout", type=float, default=None,
+                    help="stall timeout in s (timer resets on delivery); "
+                         "default max(900, steps*60)")
     ap.add_argument("--port", type=int, default=8600)
     args = ap.parse_args()
+    if args.timeout is None:
+        args.timeout = max(900.0, args.steps * 60.0)
 
     work = Work(args.packages, args.steps, args.seed, args.timeout)
     srv = ThreadingHTTPServer(("127.0.0.1", args.port), make_handler(work))

docs/features/package_courier_harness.md

@@ -79,7 +79,12 @@ The launcher boots all three processes: the opencoti-llamafile server
 (model, `-c`, `--parallel`, `--polykv-max-pools`, KV types, MTP flags —
 all pass-through), the work server, and the agent app.
 
-Defaults: 100 packages × 10 steps, 15-min stall timeout, model =
+Defaults (updated 2026-07-16, user decision + bug-2222): 100 packages ×
+10 steps; engine `--parallel 8`; pooled worker fleet driven by
+`--tps-floor 15` (fixed worker counts are only for the naive A/B leg,
+where the floor defaults to 0); work-server stall timeout auto-scales
+`max(900 s, steps × 60 s)` — the stall timer resets only on delivery,
+so long chains (e.g. 55 steps) need the scaled window; model =
 **Qwen3.6-27B-Omnimerge-v4 (bs2)** — hybrid GDN + NextN MTP, i.e. the
 serving target and the hardest pool path; everything overridable.