[ymote]

Android chat drags could wait nearly a refresh because native AChoreographer ran before Java View delivered batched input. Schedule rendering through a static Java Choreographer callback after View input, retaining no Activity and preserving the native callback-chain guard, JNI signature, and fallback for older packagers or no_android_choreographer.

Additional changes:

• Advance the first PortalList/ScrollBar fling frame immediately, then use measured frame intervals.
• Add a direct PortalList::set_capture_overload setter without script reapplication, redraw, or auto-tail reset.
• Clear stale PortalList drag samples when a child takes pointer capture. Long-press selection previously transferred capture without FingerUp, suppressing later handle movements; existing overscroll still settles.
• Stop closed Modal/Tooltip fullscreen draws, skip unused recursive GPU inventories, invalidate list topology only after retention removes entries, and avoid unchanged setter invalidations while preserving side effects.
• Blink the default TextInput caret with discrete states instead of 50 ms fades.
• Bound opt-in widget-cost attribution at 1,024 operation/type pairs. Rinx retains all 353 observed entries without overflow; presentation measurements disable profiling.

Backport official Makepad #1288 (711521dbf): normalized Android navigation with Emphasized easing after initial drawing, accurate Bezier solving, stale ShowTextIME removal, and deferred IME hiding. Adapt this to the fork's two-view pool and event APIs, and update the notes caller for width-free navigation methods.

Backport official Makepad #1296 (95e0ad5a4): shared natural-time image animation, offscreen pause, APNG/WebP compositing, compact atlases, natural-size texture metadata, and optional display-sized decoding with larger-cache reuse. Preserve asynchronous eviction recovery and error reporting. APIs absent from this fork, including image slice/crop options and cancellation scopes, are not introduced.

Main through d955125e9 is included, preserving #113 and #114. Rinx integration: hagency-org/Rinx#85. Merge with the tested commit retained in main history because Rinx pins it directly. This targets the OctoSense fork; the additional changes have not been merged into official Makepad.

Validation

Exact tested revisions: Rinx 297903d1043023d8992136a0e622fdf92df82795, Makepad 07825feab684d21f69b4afb1ec60f623c13f9183.

Test	Baseline	Candidate
Hidden native Metal: median callbacks / 8 s focused idle	207	13
Hidden native Metal: callbacks / 20 s typing/deletion	539	87
Mate 70 Air: median callbacks / 8 s focused idle	69	17
Mate 70 Air: callbacks / 20 s typing/deletion	156	48
Mate 70 Air: presentation intervals >25 ms, 12 native swipes	35/1,784 (1.96%)	26/1,781 (1.46%)

The signed HarmonyOS release packages ran on Mate 70 Air (SUP-AL90, API 24, 60 Hz). Candidate median/p95 presentation intervals were 16.622/16.665 ms, maximum 33.294 ms. The baseline includes one unresolved 1,047 ms first-interval gap (swipe setup versus app latency), retained in raw totals. These small samples do not establish a robust scrolling speedup.

macOS and HarmonyOS used an isolated 160-message Matrix fixture; typing inserted and deleted an unsent draft. The final instrument-only and native-swipe protocols passed, with no messages sent. Phone scrolling used native touch injection, profiling disabled, and RenderService timestamps for the owned surface. Native drawable captures verified rendering. Invalid USB-input and stale-history trials were excluded. Detailed captures/reports remain local; owned test sessions are closed.

Redmi Note 12 final release: long-press and handle-drag selection extends through wrapped text without moving the conversation. Focused idle recorded 1.91 presents/s versus the earlier baseline's 8.70/s. With profiling off, 24 gestures recorded 7/2,281 missed refresh opportunities (0.307%), max 33.421 ms; a same-session previous-build control recorded 7/2,252 (0.311%). No additional scrolling gain is established by that comparison. Current-head OnePlus runtime validation is not claimed.

Rinx CI passes all three checks; release unit tests record 355 passed, 0 failed, 2 existing ignores. Makepad widget tests record 292 passed and three failures also reproduced on the baseline: desktop_style::tests::mobile_typefaces_keep_symbol_fallbacks_across_appearances, widget_tree::tests::test_observe_and_find_single_node, and widget_tree::tests::test_property_patch_no_structural_rebuild. Widget compile checks pass on macOS, Android, Wasm, Linux, Windows, and OpenHarmony using installed SDKs. Existing warnings remain; iOS is untested because its Rust target is absent. Draw/platform suites passed 151/252 tests on the integrated parent; the final Makepad change is limited to PortalList. These are compile/unit results, not six-platform runtime claims.

Remaining limits

This is not a hitch-free, battery-saving, memory-leak, or native-app-parity claim. Cold room opening (about 276 ms in the earlier Android capture), safe cross-message rich-text recycling, gradual memory growth, and startup-theme persistence remain unresolved. Image-memory savings were not quantified.

Hidden macOS windows use timer pacing, so those results establish callback work, not display frame rate or physical input latency. Profiler-off scroll CPU medians were 0.99 -> 0.90 seconds in small runs under background load; no causal gain is claimed. Combined typing/scroll captures had glyph artifacts in both revisions; separate scroll-only captures rendered correctly, leaving visible-window follow-up.

The OHOS painted-frame counter remains unwired and was not used. Mobile chat children are missing from widget inspection, so native-image checks validated the composer outside timing intervals. The local fixture does not establish performance against a live encrypted server or under sustained thermal load.