IRIS_TODO.md: streaming re-lays out the whole message, from the phone's bench v2

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

docs/IRIS_TODO.md

@@ -450,3 +450,21 @@ do not duplicate it there.
   and control sizes; the emulator at two densities and the phone draw the
   same layout at the same physical size. After the bench setup is
   finished, before P1 draws any new screen.
+
+## From the phone, bench v2 (2026-09-06): streaming re-lays out the whole message
+
+- [ ] **Streaming a delta into a long message costs a full text layout of
+  that message.** Iris's phone report (`docs/bench/iris-phone-v2-2026-09-06.md`):
+  the stream phase is the one place iris is behind Compose (p50 18.2 ms vs
+  13.4 ms; p99 level at ~43 ms). `TranscriptScreen::apply` replaces only
+  the last row, but that row is the growing message, and replacing it
+  re-renders its markdown and re-shapes the entire paragraph run through
+  parley on every event. Compose pays a reparse (8.6 ms mean) for the
+  same event. What "done" looks like: a streamed delta re-lays out only
+  the block it lands in (the last paragraph or code block), with earlier
+  blocks' layouts kept -- which needs a row to be a column of per-block
+  `Text`s rather than one `TextEdit` for the whole message, or parley's
+  layout to be split at block boundaries; measured by the stream phase's
+  p50 dropping below Compose's on the phone. Do this after the four bench
+  v2 defects (stale primitives, finger fling, decay curve, IME show) are
+  closed, since they are what make the run unrepresentative today.

docs/bench/iris-phone-v2-2026-09-06.md

@@ -0,0 +1,58 @@
+# iris bench v2 report from Iris's phone, 2026-09-06
+
+Build 2e3f4ad (bench v2, fling physics, keyboard-wipe fix, dp unit), run
+by Iris on her Pixel 9 Pro XL, verbatim. The display was at **120 Hz**
+(8.3 ms budget) where `compose-phone-v2-2026-09-06.md` ran at 60 Hz, so
+compare the millisecond percentiles, not `late`.
+
+Side by side (Compose 60 Hz / iris 120 Hz, p50 / p90 / p99 ms): fling
+5.5/8.7/11.6 vs 3.8/6.9/12.6; stream 13.4/31.7/42.5 vs 18.2/35.8/43.1;
+type 7.3/13.2/16.5 vs 7.2/9.2/11.2; keyboard: iris could not show the IME
+(phase invalid). Process CPU 69.6 s over 125 s vs 40.6 s over 150 s; peak
+RSS 577 MB vs 379 MB; battery current mean 571 mA vs 452 mA.
+
+Iris's observations on the same run: "the scrolling is not similar at
+all. It does not fling for me yet [with a finger], and the test also seems
+to give it a constant velocity and abruptly stop it at some point. Also
+unsure what's going on in that image with the compaction" -- her
+screenshot shows the `Compacted: 180000 -> 20000 tokens.` row drawn twice
+overlapping, and once more below the composer bar: primitives of a
+replaced/removed row surviving in the GPU buffers, the same shape as the
+header drawn twice after a keyboard resize.
+
+```
+iris bench report
+per phase:
+  fling: 1783 frames over 53.2s
+    late: 104 (5.8%)
+      total  p50 3.8ms  p90 6.9ms  p99 12.6ms
+    worst 29.1ms
+  stream: 401 frames over 21.3s
+    late: 306 (76.3%)
+      total  p50 18.2ms  p90 35.8ms  p99 43.1ms
+    worst 43.8ms
+  type: 1202 frames over 65.7s
+    late: 309 (25.7%)
+      total  p50 7.2ms  p90 9.2ms  p99 11.2ms
+    worst 15.3ms
+  keyboard: 9 frames over 9.7s
+    late: 9 (100.0%)
+      total  p50 12.0ms  p90 12.9ms  p99 12.9ms
+    worst 12.9ms
+
+frames:
+  3395 frames over 149.9s at 120Hz (8.3ms budget)
+  late: 728 (21.4%)
+  total  p50 5.0ms  p90 10.9ms  p99 36.6ms
+  worst 43.8ms
+  cpu_p50 2.0ms  gpu_wait_p50 2.6ms
+
+bench:
+  fling: 8 flings out + 8 back at 12000px/s, travel start=idx=651/off=1217px outward=idx=651/off=101536px end=idx=651/off=1022px
+  scroll: 6 cycles (24 swipes, legacy tween), streamed 400/400 fixture events
+  type: 600 characters inserted then deleted, one per 50ms
+  keyboard: could not be shown (5 attempts, 0 confirmed visible)
+  process CPU time over this run: 40603ms
+  peak RSS: 379156kB
+  battery current: mean -452353µA over 149 samples (min -1753125, max -204687)
+```