Publish settled-layout iris benchmark

This commit is contained in:
iris committed 2026-09-09 16:57:00 -04:00
1 parent 6d813942d0
commit c1bc5883c1
2 files changed
+19 -1

No files matched your search

+19 -1
View File
@@ -18,7 +18,25 @@ to the benchmark card.
The rest of this file is the runbook that was delivered with the APKs.
**2026-09-09 (latest): pinned growth moves the retained transcript once.**
**2026-09-09 (latest): growing layout settles in the frame that changed it.**
`iris-bench-arm64.apk` rebuilt from ai-app-2 `5ece49b`. This fixes the two
regressions found in the retained-transcript build: a newly streamed markdown
block no longer spends its first frame beside/overlapping the previous block,
and the multiline composer grows to its final viewport before it is masked, so
the first visible line and its padding are not clipped.
The fix is entirely in iris. A changed child's size was already measured up
through its ancestors; iris now lays that changed branch back down through the
final boxes in the same frame, and `Sized` no longer accepts a provisional box
for its descendants as final. The dedicated LazySpan regression and the real
phone-sized composer regression both fail without their respective fixes.
The retained move remains intact: the 401-event upload rig reports instance
uploads at **1.1% against a 1.1% floor** (p50 **1,488 bytes**) and an exact
11,569-slot arena for 11,569 live primitives. Streamed-frame CPU is p50 0.12ms,
p99 0.43ms, with no frame over 8.3ms in the layer-1 release profile.
**2026-09-09: pinned growth moves the retained transcript once.**
`iris-bench-arm64.apk` rebuilt from ai-app-2 `fffed42`. This is the
framework-level `LazySpan` change: when a streaming message grows and pushes
the transcript upward, retained rows keep stable local boxes and one shared