Record final deck geometry performance comparison and raster limits

This commit is contained in:
2026-11-07 19:55:24 +00:00
parent 17569034e0
commit b82e89f66f
+44 -1
View File
@@ -29,7 +29,7 @@ helper now uses the mathematically equivalent affine transform for these cards,
while actual Flip/Rolodex rotations keep perspective. Regression verifies the
same projected points instead of requiring an unnecessary perspective matrix.
## Quiet before/after confirmation
## Intermediate geometry comparison (superseded by final comparison below)
Two repetitions × six warmed touch switches per build, same fixture and 850 ms
cadence. Root paused other builds/tests/captures for this comparison. Before is
@@ -69,3 +69,46 @@ Tests cover no repeated child-constraint layout as painted height changes,
fixed occupied chrome during a fade, actual resting peek height, retained
chat drafts for every style and equivalent affine projection. Integration
browser review remains with the parent task.
## Final release geometry comparison
Repeated the quiet pair after the final deck polish: the incoming card is already
full-height behind the outgoing opaque front, avoiding a blank gap at the top.
Both builds use exactly that same geometry and opaque surface fill (3daf9ce0).
Only the fixed viewport/chrome layout and affine flat-depth optimization differ.
Two repetitions × six warmed switches each, same rich fixture and 850 ms cadence;
root paused other builds/tests/captures. These final results supersede the earlier
geometry's apparent 25% rAF-tail improvement.
| Metric | Cost fixes disabled | Cost fixes enabled |
| --- | ---: | ---: |
| Flutter build p50 | 9.0 ms | 1.6 ms |
| Flutter build p95 | 39.7 ms | 20.6 ms |
| Flutter build sum | 895.4 ms | 502.8 ms |
| Total frame p50 | 45.7 ms | 39.5 ms |
| Total frame p95 | 79.2 ms | 78.8 ms |
| Raster p95 | 58.4 ms | 71.0 ms |
| Browser rAF interval p95 | 66.7 ms | 66.7 ms |
| Long tasks (>50 ms) | 36 | 23 |
| Flutter frames recorded | 86 | 96 |
The layout fix reduces typical build work by about 82%, and typical total frame
cost by about 14%. **Raster tail worsens, rAF p95 does not improve, and total-frame
p95 is essentially unchanged.** Fixed full-size content behind a changing paint
clip still has renderer/filter cost; this patch does not solve that cost or
prove smooth native-device FPS. The prior quarter-improvement cadence result
must not be presented as the final release result.
Final reports/raw timings/CPU profiles: /tmp/mobile-motion-final-before/ and
/tmp/mobile-motion-final-after/. Both report no browser errors and all 12
selections confirmed. The disabled-cost build used the same widget structure
with fixed sizes disabled, progressive occupied chrome restored and the previous
perspective transform path; source was restored byte-for-byte after building.
All private profiling browser/server processes exited. No Android hardware test
was possible. Parent's final browser capture covers the final geometry and
verifies the resting peek reaches its true content bottom.
Final performance regression checks: 61 workspace/motion tests passed
(/tmp/memwatch/out.Yvac5n), analyzer clean (flutter-analyze.wtLee1). The subsequent
geometry change has its separate parent/animation-lane test evidence.