Stop Android navigation from rebuilding the whole screen (MOB-146)

[GenericJam]

Stacked on #57 (the frame timing that makes this measurable).

Screens rendered through AnimatedContent(targetState = state, contentKey = { it.navKey }). AnimatedContent wraps each content in key(contentKey), so every push, pop and reset disposed the outgoing composition and built the incoming one from nothing — structurally what .id(currentNavVersion) did on iOS before MOB-129.

Now one mount point at a fixed structural position. That position is its identity, no key is used, and a navigation replaces the tree inside it. The slide is driven by an animated layer translation rather than an enter/exit transition, because those only fire on insert/remove and insert/remove is what cost the time.

Measured

Physical moto g power, 1600-node screen, baseline and fix back to back in one session, n=20/10/10:

transition	before	after
none	227 ms	128 ms	44% faster
push	794 ms	374 ms	53% faster
pop	784 ms	440 ms	44% faster

What this does not do, since an earlier draft of this work claimed otherwise: it does not make navigation as cheap as a re-render. Nothing in a Mob screen can be skipped — MobNode holds a Map and a List, so Compose infers it unstable, and this toolchain has no strong skipping — so every recomposition walks all 1600 nodes either way. What is removed is the dispose and the slot-table teardown and rebuild. Navigation still costs ~3x a re-render, before and after. GenericJam/mob#TBD-162 covers the rest.

Those earlier numbers (161ms push) were a measurement artifact: the tree was installed by a LaunchedEffect, which runs after composition, so the frame native_stats measured still showed the old tree. Rendering straight from state fixed the artifact and a blank-screen bug together.

iOS's two-slot presentation was implemented first, and removed

Parking the outgoing tree buys depth-1 retention there. Here a parked Compose subtree recomposes on every render of the active screen (measured: 6 recompositions of the parked node across 6 re-renders), taking a steady-state re-render from 151ms to 273ms — paying on the common path to save on the rare one. Worth revisiting under MOB-162, since the cause is the same unstable MobNode.

Every navigation re-keys the frame trackers

The gate has each tracked node remember the generation current when it composed and refuses older writes. AnimatedContent used to make the incoming tree a fresh composition, so that capture was renewed for free. With the mount point preserved, nodes Compose reuses would keep a superseded generation and their frame writes would be refused for ever — element_frames silently losing ids, tap_id no longer finding them.

navKey is the epoch: it moves under the same if (transition != "none") in setRootJson that bumps the generation, so the two cannot drift apart. Verified on device across five consecutive navigations.

Two silent bugs found and fixed on the way

The app rendered blank. The animation was driven from LaunchedEffect(state); that cancels on key change and state is new every render, so any re-render landing during the 300ms slide cancelled animateTo and froze the offset off-canvas. The BEAM went on reporting the correct screen and assigns throughout — no probe an agent has would have caught it. Now keyed on navKey.

A black wedge on light themes. Only one screen is mounted during a slide and the window background beneath is hardcoded black in styles.xml. The container now paints MaterialTheme.colorScheme.background.

Also gone with the enter/exit transitions, and documented: a reset no longer cross-fades (the default for Mob.Socket.reset_to/4), and a push no longer parallaxes the outgoing screen.

Tests

The identity change is gone by both known routes (AnimatedContent and a bare key(state.navKey)), the slide parks before animating, the animation is keyed on navKey not state, and RenderNode sits inside the epoch provider — a provider wrapping an empty body passes a presence check while every tracked node reads the default epoch for ever. Mutation-checked.

Suite 418/420 (2 pre-existing MOB_DIR env failures), credo and ktlint clean.

🤖 Generated with Claude Code