docs/bench: the Compose P0 report from Iris's phone, verbatim

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

docs/bench/compose-phone-2026-09-06.md

@@ -0,0 +1,73 @@
+# Compose bench report from Iris's phone, 2026-09-06
+
+The Compose half of P0 (RUST.md), run by Iris on her own phone and pasted
+back verbatim. The iris half's report goes beside it in this directory
+when it exists. Her caveat, worth keeping with the numbers: "I don't think
+this is entirely fair because the UI for iris is more minimal" -- the
+Compose screen also draws the usage bar, the status row and tool cards,
+which the iris bench screen does not yet. Her impression of the iris build
+before its first-touch bug: "it already feels very smooth so far".
+
+What to read first: the phone runs at 120 Hz, so the budget is 8.3 ms;
+`late` is measured against that. Compose's tail is the streaming phase --
+`markdown reparsed while streaming: 396, 8.5ms mean, 25.8ms worst` and
+`record: one block: 398, 6.3ms mean, 19.9ms worst` -- which is exactly the
+path iris's `TranscriptScreen::apply` (replace the last row only) is meant
+to beat. Process CPU over the run is 20.9 s of a 38.5 s run; peak RSS
+587 MB; battery current mean 419 mA.
+
+```
+ai-app render report
+device: Pixel 9 Pro XL (Google), Android 17
+build: release
+
+transcript:
+  43 events, 41 rows, 93 units loaded
+  viewport 1333px, 2 units visible
+  on screen: the list's own 0px, AssistantMsg 24520px
+  0 tool calls and 0 groups open
+
+frames:
+  1613 frames over 38.5s at 120Hz (8.3ms budget)
+  late: 742 (46.0%)
+  total  p50 7.7ms  p90 29.2ms  p99 41.1ms
+  waited p50 0.5ms  p90 12.9ms  p99 27.2ms
+  input  p50 0.0ms  p90 0.0ms  p99 0.0ms
+  anim   p50 1.1ms  p90 5.8ms  p99 9.5ms
+  layout p50 0.0ms  p90 0.1ms  p99 0.2ms
+  draw   p50 0.4ms  p90 15.9ms  p99 27.9ms
+  sync   p50 0.1ms  p90 0.5ms  p99 1.0ms
+  issue  p50 1.1ms  p90 1.7ms  p99 3.0ms
+  swap   p50 0.4ms  p90 0.5ms  p99 0.7ms
+  gpu    p50 1.8ms  p90 2.1ms  p99 6.6ms
+
+where the draw phase went:
+  draw phase 3.83ms per frame, of which:
+    the transcript: 0.25ms (measure 0.15, place 0.10, record 0.00)
+    everything else: 3.58ms (93%)
+
+work since this was last copied:
+  draw: the whole transcript: 12, 0.2ms total, 0.0ms mean, 0.0ms worst
+  grouped tool runs: 398, 10.3ms total, 0.0ms mean, 0.1ms worst
+  markdown cut into pieces: 1, 0.0ms total, 0.0ms mean, 0.0ms worst
+  markdown parsed while composing: 7, 3.0ms total, 0.4ms mean, 0.5ms worst
+  markdown ready: 46
+  markdown reparsed while streaming: 396, 3384.6ms total, 8.5ms mean, 25.8ms worst
+  markdown warmed: 1, 1.4ms total, 1.4ms mean, 1.4ms worst
+  measure: the whole transcript: 957, 248.7ms total, 0.3ms mean, 15.7ms worst
+  message composed: 403
+  message cut into parts: 1, 0.1ms total, 0.1ms mean, 0.1ms worst
+  place: the whole transcript: 1319, 162.4ms total, 0.1ms mean, 1.3ms worst
+  record: one block: 398, 2498.7ms total, 6.3ms mean, 19.9ms worst
+  session screen recomposed: 413
+  status row recomposed: 1
+  unit composed: 538
+  units flattened: 399, 44.3ms total, 0.1ms mean, 0.5ms worst
+  usage bar recomposed: 413
+
+bench:
+  scroll: 6 cycles (24 swipes), streamed 400/400 fixture events
+  process CPU time over this run: 20907ms
+  peak RSS: 587356kB
+  battery current: mean -418509µA over 39 samples (min -1988281, max -107812)
+```