perf: Write executeMount int buffer to Java in a single JNI call

[shubh73]

Summary:

FabricMountingManager::executeMount wrote the int buffer to Java with one SetIntArrayRegion per writeInt/writeIntArray, which is ~4.5k JNI calls for a commit that mounts 1500 views. The exact size is already known from computeBufferSizes, so this fills a std::vector<jint> instead and copies it once before createIntBufferBatchMountItem. The int[] Java receives is identical.

GetPrimitiveArrayCritical isn't an option here since the object writes call into Java between int writes, and on ART GetIntArrayElements copies movable arrays anyway.

Time spent in executeMount for one commit on an OPPO CPH2667 (Dimensity 6300, Android 16), release build. Each cell is the median over 114 commits from 6 interleaved runs:

Commit	JNI calls	Before	After	Change
Remove 1500 views	3004 → 1	2.14 ms	1.88 ms	−12%
Add 1500 views	4505 → 1	4.06 ms	3.86 ms	−5%
Restyle 1500 views (size, padding, color)	4505 → 1	5.25 ms	5.04 ms	−4%

The slowest commits improve more (e.g. 19.6 → 6.4 ms when restyling). The int transfer on its own goes from 174 → 87 µs per 1500-view commit. On an Apple silicon emulator JNI calls are ~5x cheaper and the difference is within noise.

Changelog:

[ANDROID] [CHANGED] - Write Fabric mount instruction ints to Java in a single JNI call

Test Plan:

• Built a variant that also fills a shadow int[] the old way and compares it with the array sent to Java on every commit: 1390 commits on device and 782 on emulator across Playground (1500 views), FlatList, Text and Image, 0 mismatches.
• RNTester (yarn android and release) on device and emulator: Playground, FlatList, Text and Image render as before.
• yarn lint, yarn flow-check, yarn format-check-cpp pass.
• Couldn't run FabricMountingManagerTest / FabricMountingManagerInstrumentationTest, they're Buck-only.

🤖 Generated with Claude Code