docs: record final phone benchmark comparison

docs/bench/iris-phone-v3-2026-09-09.md

@@ -0,0 +1,76 @@
+# iris phone bench v3, 2026-09-09
+
+Fresh release run on Iris's Pixel 9 Pro XL after the frame-report corrections,
+realistic streamed blocks, retained arena uploads and fling-clock fix. Iris
+reported that everything was visually working as intended. This run is paired
+with `compose-phone-v3-2026-09-09.md`.
+
+The iris reporter resolved the phone at **120 Hz** (8.3 ms budget). The paired
+Compose report says 60 Hz, but Iris observed what looked like 120 Hz and its
+fling count works out to about 118 fps after the benchmark's deliberate rests
+are removed. Compose nevertheless calculated `late` using its reported 60 Hz
+and 16.7 ms budget, so the two apps' `late` percentages are not directly
+comparable. Against the real 8.3 ms budget, iris recorded 0.5% late frames
+overall. Streaming had no late frames or missed vsyncs, with p50 3.1 ms and
+p99 5.0 ms, against Compose's p50 13.8 ms and p99 42.6 ms. The five missed
+vsyncs in the fling phase are 0.15% of its 3,409 frames; Iris could not see lag
+in the finished UI.
+
+Across the similarly long complete runs, iris used 37.5 seconds of process CPU
+against Compose's 70.5 seconds (47% less) and peaked at 382,796 kB RSS against
+581,084 kB (34% less). Mean battery current was 226 mA against Compose's 591
+mA; that is encouraging but remains a device-state-sensitive reading rather
+than a controlled energy measurement.
+
+The keyboard line does **not** describe a failed opening. The type phase opens
+the IME and leaves it visible. The keyboard phase's first `show_ime` is
+therefore a no-op rather than a new hidden-to-visible transition; it then
+observes five hides and four reopens. Including the opening in the type phase,
+the complete run observed five openings and five closings.
+
+`acquire` is the wait for a swapchain image, not iris's own work. In particular,
+the fling's 8.0 ms total median consists mostly of a healthy 6.0 ms acquire
+wait; its median build work is 0.4 ms. `late` is judged on work with that
+acquire wait removed.
+
+```text
+iris bench report
+input/frame trace: off
+per phase:
+  fling: 3409 frames over 33.2s
+    late: 11 (0.3%)  missed vsyncs: 5
+      total  p50 8.0ms  p90 9.0ms  p99 10.5ms
+      build p50 0.4ms  acquire p50 6.0ms  submit p50 1.6ms
+    worst 17.4ms
+  stream: 401 frames over 21.3s
+    late: 0 (0.0%)  missed vsyncs: 0
+      total  p50 3.1ms  p90 4.0ms  p99 5.0ms
+      build p50 1.1ms  acquire p50 0.3ms  submit p50 1.6ms
+    worst 7.0ms
+  type: 1240 frames over 64.5s
+    late: 14 (1.1%)  missed vsyncs: 0
+      total  p50 5.6ms  p90 6.8ms  p99 9.1ms
+      build p50 2.5ms  acquire p50 0.4ms  submit p50 2.6ms
+    worst 12.8ms
+  keyboard: 341 frames over 9.8s
+    late: 0 (0.0%)  missed vsyncs: 0
+      total  p50 2.1ms  p90 2.8ms  p99 4.5ms
+      build p50 0.3ms  acquire p50 0.2ms  submit p50 1.5ms
+    worst 8.5ms
+
+frames:
+  5391 frames over 128.8s at 120Hz (as the display reports it; drew 120fps) (8.3ms budget)
+  late: 25 (0.5%)
+  total  p50 7.5ms  p90 8.8ms  p99 10.2ms
+  worst 17.4ms
+  build_p50 0.4ms  acquire_p50 5.3ms  submit_p50 1.7ms
+
+bench:
+  fling: 8 flings out + 8 back at 12000px/s, travel start=idx=653/off=1557px outward=idx=528/off=909px end=idx=650/off=-103px ticked=frame-loop
+  scroll: 6 cycles (24 swipes, legacy tween), streamed 400/400 fixture events
+  type: 600 characters inserted then deleted, one per 50ms
+  keyboard: shown 4/5, hidden 5/5 (confirmed via on_insets_changed)
+  process CPU time over this run: 37457ms
+  peak RSS: 382796kB
+  battery current: mean -226354µA over 128 samples (min -404687, max -107812)
+```