Tao: track upstream CMP-10678 (RectManager EDT dispatch) and remove RectManagerEdtGuard once fixed

[kdroidFilter]

Follow-up to #551 / PR #554.

RectManagerEdtGuard closes the Compose 1.12 RectManager EDT race for every app that does not register debounced rect callbacks. A documented residual remains: an app explicitly using Modifier.onLayoutRectChanged / onFirstVisible / onVisibilityChanged (with debounce) can still get one EDT dispatch if a layout pass runs >16 ms past a mid-pass re-arm with no layout/semantics notification in between — because a real (non-pinned) deadline lowers minDebounceDeadline and the pass-end scheduleDebounceCallback arms a live EDT timer that the guard can only cancel at the next observable hook.

Proposed full closure (opt-in, layered)

Swap RectManager.executeDelayed (private field) for a scene-thread implementation at owner creation:

• The ExecuteDelayed implementation itself can be static (@Suppress("INVISIBLE_REFERENCE"), no dynamic Proxy) — it schedules delay(millis) on the scene's coroutine context, exactly like the guard's post-frame wakeup.
• The only reflective operation is a single Field.set on RectManager.executeDelayed.
• GraalVM native-image: needs one reachability-metadata field entry for androidx.compose.ui.spatial.RectManager; must verify on a real runGraalvmNative build that a reflective write to a final instance field works with the metadata (final-field writes are the sensitive part).
• Layered behavior: attempt the swap at first owner; on success the residual disappears entirely; on any failure (missing metadata, future Compose renames the field) log one warning and keep the current guard — never worse than today.

Exit criterion

This issue becomes obsolete if JetBrains fixes the upstream bug (desktop postDelayed runs on skiko's hardcoded MainUIDispatcher/EDT — their code carries TODO CMP-7153; a standalone no-framework reproducer crashing 3/3 off-EDT and never on the EDT has been prepared for the YouTrack report). Once the upstream fix ships, both the guard and this follow-up can be removed for that Compose version.

[kdroidFilter]

Deeper analysis + empirical testing show the residual is unreachable on Compose 1.12.0-rc01, so the reflective layer-2 closure is not currently needed.

Two mechanisms combine to make the guard's pin complete even for apps registering debounced rect callbacks:

1. ThrottledCallbacks.triggerDebounced early-returns (if (minDebounceDeadline > currentMillis) return) while the deadline is pinned, and it is the only place that recomputes the deadline from the registered entries.
2. The mid-pass lowering branch in ThrottledCallbacks.fireWithUpdatedRect / fire is an upstream no-op: if (currentMinDeadline > 0 && thisDeadline < currentMinDeadline) { minDebounceDeadline = currentMinDeadline } assigns the old value back (thisDeadline was clearly intended).

So with the guard installed, minDebounceDeadline only ever becomes real inside the guard's own post-frame hook, on the scene thread, with the dispatch cancelled microseconds later — no live EDT timer can ever arm.

Verified empirically with examples/rect-stress-demo (added as a standing sentinel): throttled+debounced onLayoutRectChanged entry moving every frame, LazyColumn mid-pass dispatches, >16ms draw and placement phases, periodic remounts, and a wrong-thread detector in the callback (exit 55 with proof if RectManager ever invokes it off the scene thread). No detection, no crash.

Keeping this issue open but re-scoped: (a) onScenePulse stays as defense-in-depth, and the sentinel demo must keep passing, because if upstream fixes the no-op assignment the lowering path comes back; (b) revisit when the upstream YouTrack report gets a fix version — the guard (and this issue) can be deleted entirely once RectManager schedules on the owner's context.

[kdroidFilter]

Upstream report filed: CMP-10678 — Desktop: RectManager delayed dispatch runs on the AWT EDT and corrupts RectList when a ComposeScene is driven from a non-EDT thread. Exit criterion for this issue (and for removing RectManagerEdtGuard) is the fix version of CMP-10678. Note for whoever picks this up: if the upstream fix lands as a correction of the no-op deadline-lowering assignment in ThrottledCallbacks (rather than rescheduling postDelayed off the EDT), the lowering path comes back and the guard's onScenePulse/afterFrame layers become load-bearing — the rect-stress-demo sentinel covers that scenario.