android: Clear and describe Java exceptions in the core JNI calls

[Bombatomica64]

Part of #31 (PR 2 of the proposed list). This PR changes only the core files under src/android/. The platform packages are left for a follow-up (PR 6 in #31).

Problem

When a call into Java throws, the exception stays pending. jni 0.22 then refuses every later JNI call in the same frame. Until with_env returns, they all fail with the same bare Java exception was thrown, and at the end with_env replaces the closure's result with that exception. In the core files:

• Some calls convert the error with map_err(|e| e.to_string()) and leave the exception pending: the IME show/hide calls, the keyboard layout and thermal queries, and update_ime_position.
• Others clear it by hand, so its message is lost (find_app_class, set_system_chrome, the platform views). Some also drop the error with let _ =.
• hide_keyboard_android and reset_keyboard_composition clear the exception before converting the error, so the message is lost there too.
• set_system_chrome falls back to the pre-API-30 setSystemUiVisibility flags on any failure of getInsetsController, not only when the method doesn't exist.
• get_string reads any object as a java.lang.String without checking its type (nativeIme passes it whatever Java sent).

Change

• New JniResultExt::or_clear(env), for jni::errors::Result. On error it clears the pending exception and returns its toString(), so errors read like java.lang.SecurityException: <message> instead of Java exception was thrown.
• New JniResultExt::or_catch(env, class), for an exception that is an expected result: an instance of class is cleared and gives Ok(None). Any other exception is handled like or_clear. set_system_chrome uses it so that only java/lang/NoSuchMethodError (API < 30) leads to the fallback.
• take_exception(env) clears the pending exception and returns its description. set_host_activity uses it in place of its own clear.
• The calls into Java in jni.rs, host.rs, platform.rs, window.rs, platform_view.rs and document_picker.rs now use these. Their errors are returned, or logged where the function returns nothing (warn; debug for the per-keystroke IME calls).
• get_string checks is_instance_of(java/lang/String) first. For an object of another type it logs and returns an empty string.
• The existing JniExt::e() is unchanged. Its doc now says that it leaves the exception pending, so it is for JNI calls that don't call into Java.

accessibility.rs is not changed: it works through accesskit_android's jni 0.21 and already clears its exceptions. The JNI exports (nativeIme, nativeOnDeepLink, …) are not changed either: #31's PR 1 edits the same lines, and this keeps the two PRs independent.

Testing

Run on this branch with the Android NDK 27, target aarch64-linux-android:

• cargo fmt --all -- --check
• cargo check --target aarch64-linux-android --all-targets, plus cargo check of example/
• cargo clippy --target aarch64-linux-android: the same 8 warnings as on main, none new
• cargo test --lib: 46 passed
• Merges cleanly with the docs PR from Proposal: rework the Android JNI layer (Activity lifetime, exceptions, UI thread), as small PRs #31 (PR 1).

An earlier version of these changes ran on a OnePlus CPH2581 (Android 16), through the host entry point, as part of the full rework in Bombatomica64#1. This PR is that work ported onto the current main: #24–#30 moved the code (per-host Activities, night mode, assets), so the hunks were redone by hand rather than cherry-picked.

Follow-ups

The same helpers in the platform packages (PR 6 in #31), and the host Activity registry (PR 3), which uses or_clear.

🤖 Generated with Claude Code