Time the native half of an Android frame (MOB-146)

[GenericJam]

Mob.RenderStats could time the BEAM half of a render but not the native half, and on a dense screen the native half is most of the cost. On Android it could see none of it — there was no native implementation at all, so native_summary/1 returned {:error, :unsupported}, which is indistinguishable from "you forgot to enable it."

MOB-146 argues Android navigation disposes and recreates the composition, the same defect MOB-129 fixed on iOS. Without a number, that fix lands with no before and no after. This is the prerequisite the ticket asks for.

The bracket is the whole problem

Both obvious closing brackets are wrong, and neither fails — they return a plausible number. Documented in place so nobody re-derives them:

• MessageQueue.IdleHandler is the literal analogue of the CFRunLoopObserver(.beforeWaiting) iOS uses. Compose requests its frame through Choreographer, and a vsync callback arrives asynchronously rather than sitting in the queue — so between request and vsync the queue is genuinely empty. The handler fires there, before any measured work, and reports the cost of a field write.
• Registering from inside a posted Runnable fails less visibly. ViewRootImpl.scheduleTraversals installs a sync barrier holding non-async messages until doTraversal. With a traversal already pending, the post is held while Compose recomposes at frame V, so it registers for V+1.

That second one was caught in review as a suspicion and confirmed by measurement: removing the outer post dropped a one-field re-render from 266ms → 221ms and a pop from 916ms → 788ms. The bracket had been absorbing roughly an extra frame, biased upward exactly under the load being characterised.

So the callback is registered straight from the calling thread against a Choreographer captured on main at init, then a post from inside it closes the bracket after the synchronous traversal.

Baseline

Physical moto g power, 1600-node screen:

transition	n	p50
none (one field)	10	221 ms
push	4	818 ms
pop	4	788 ms

Navigation costs 3.7x a re-render of the same tree.

Corroborated before it was believed — in the same window the platform logged Davey! duration=1414ms and Choreographer: Skipped 67 frames (~1120ms). The measured figures sit just below, which is right for a bracket closing before GPU swap.

Review fixes

• Clear on enable only. Clearing on disable destroyed the window the caller was about to read — native_disable/0 documents that samples stay readable, and the natural shape is enable, drive, disable, read. It failed by returning an empty summary, which looks like the feature being off.
• Escape transition. It arrives from nif_set_transition, which takes any atom up to 15 chars verbatim; one containing a quote costs the reader the whole window, not one sample. iOS gets this free from NSJSONSerialization.
• System.nanoTime rather than elapsedRealtimeNanos, which counts deep sleep — a run spanning a screen-off would report minutes.
• Start after MobJson.parseNode.setRootJson runs synchronously from the NIF, so set_root_us already spans the parse; measuring it here double-counts.
• Cache the main Handler instead of allocating per sample.

Tests

Both halves exist and are @JvmStatic; the emitted keys are the ones Mob.RenderStats parses; the bracket rides a frame callback registered without an intervening post; the clock ignores deep sleep; enabling clears and disabling does not; the transition is escaped.

Comments are stripped before matching, per the convention in native_frame_stats_test.exs — this file names its rejected alternatives, so a plain substring search would find them in the prose explaining why they aren't used.

Suite 413/415 (2 pre-existing MOB_DIR env failures), ktlint and credo clean.

Requires GenericJam/mob#136. MobBridge.kt is generated once and never re-rendered, so existing apps must be regenerated.

🤖 Generated with Claude Code