Let Mob.RenderStats see the native half of an Android frame (MOB-146)

[GenericJam]

native_summary/1 returned {:error, :unsupported} on Android because there was no native implementation — which looks identical to "you forgot to enable it." Adds native_stats/0 and native_stats_enable/1 there. The Erlang stubs and all of Mob.RenderStats were already platform-agnostic, so only the native half was missing.

The buffer lives in the app's generated MobBridge.kt rather than here: the measurement can only be taken on the main thread, so keeping it beside the writer avoids a JNI hop per sample. These NIFs fetch and toggle it the way nif_element_frames fetches frames, and both Kotlin methods are cacheOptional so an app generated before they existed still loads.

First Android navigation baseline

Physical moto g power, 1600-node screen: none 221ms p50, push 818ms, pop 788ms. Navigation costs 3.7x a re-render of the same tree — the dispose-and-recreate gap MOB-146 exists to close, now a number rather than an inference.

Corroborated against the platform's own accounting before it was believed: Davey! duration=1414ms and Choreographer: Skipped 67 frames in the same window.

Android's apply_us is not iOS's apply_us

Stated plainly in the decision record, because a differential test (MOB-157) would otherwise compare them and conclude something false. iOS starts its clock on the main thread after the dispatch hop; Android starts on the NIF thread, so its interval includes a thread handoff and up to one vsync of queue latency. It is a before-and-after tool for one platform, not a cross-platform comparison.

Review fixes

• native_call/3 now converts a returned{:error, :not_loaded} to :unsupported. Android signals a missing bridge method by returning rather than raising, so it took the one path that wasn't converted — leaking a fourth shape past a @spec promising two, and every app generated before this release hits it.
• Exact atom compare in native_stats_enable rather than a four-byte prefix, which accepted :truthy and :true_x as "on" — the typo-enables-it case the comment beside it claimed to refuse. Anything that is neither true nor false is now badarg, matching iOS.
• native_stats added to iOS's caps[] and Mob.Test.probe_keys/0. An existing test enforces three-way agreement between the two platforms and Elixir, and caught this — a key advertised by one native half and not the other. Good test.
• Moved a comment that had landed above sample_region's value instead of render_stats', inside the exact positional array it describes.

Tests

The capabilities test now checks the arrays pairwise rather than by length. Counting cannot catch the failure it exists to catch: inserting a key mid-array while appending its value at the end keeps the counts equal and produces exactly the misordering, where each capability reports a different NIF's availability. Mutation-checked by swapping two values.

Also pins the Android JSON shape against Mob.RenderStats: a wrapped ring (recorded 250, retained 240, dropped 10), an escaped quote in a transition, and an unescaped one — the last asserting it costs the whole window rather than one sample, so the escaping in MobBridge.kt reads as load-bearing rather than defensive tidiness.

Suite 1545 passed, credo clean.

Pairs with GenericJam/mob_new#57.

🤖 Generated with Claude Code