fix(android): ship the AAR's native libraries with their debug info

[andiwand]

A consuming app's ndk.debugSymbolLevel has only the merged native libraries to extract from, and both halves of this repo were throwing that away. Play reported every frame of a crash inside the core as a bare address.

Two strippers, not one

build_native.py stripped with --strip-unneeded. That was the obvious culprit, and fixing it alone does nothing — AGP's StripDebugSymbolsTask runs --strip-unneeded again on the way into the AAR.

The first push of this PR proved it: build_native.py produced a 7,326 KiB libodr_jni.so, and the AAR CI built from it contained 5,760,752 bytes — byte-identical to the fully stripped one. Caught in review; the evidence is that run's AAR artifact.

So this needs both: the script leaves the libraries alone, and packaging.jniLibs.keepDebugSymbols holds AGP off them.

Size, measured

libodr_jni.so per ABI — the NDK compiles with -g in every configuration:

ABI	unstripped	--strip-debug	--strip-unneeded (before)
arm64-v8a	72.1 MiB	7.15 MiB	5.49 MiB
armeabi-v7a	60.8 MiB	5.42 MiB	3.89 MiB
x86	59.6 MiB	6.40 MiB	5.17 MiB
x86_64	69.1 MiB	6.97 MiB	5.64 MiB

With libc++_shared.so that is 292 MiB uncompressed across four ABIs — but DWARF deflates about 4:1, so the published AAR is 76.7 MiB (80,459,111 bytes), against ~7 MiB before.

Verified end to end

• The AAR CI now builds carries the debug info: arm64-v8a/libodr_jni.so is 75,638,648 bytes, up from 5,760,752
• Both emulator matrices (API 26 and 36) pass against the unstripped libraries
• Test-published as 6.1.1-test2 — the central portal accepted the ~77 MiB bundle and listed the AAR at exactly the 80,459,111 bytes CI produced. Deployment dropped afterwards.

Why the size is the right trade

Debug symbols never reach a phone. AGP puts them in BUNDLE-METADATA/com.android.tools.build.debugsymbols/ inside the .aab; play uses them to symbolicate and serves device APKs built from the stripped libraries. The cost is maven central storage and developer builds on a cold cache — not app size.

The alternative was --strip-debug, keeping .symtab only: ~32 MiB uncompressed, function names but no line numbers and inlined frames collapsed into their callers. Since end users pay nothing either way, the full DWARF is worth it — and it is what OpenDocument.droid already had via conan, before opendocument-app/OpenDocument.droid#558 moved it onto this AAR.

Follow-up worth its own change

publishToMavenCentral uploads and exits without waiting for validation, so a release cannot tell a validated deployment from a failed one. The maven jar's path polls and reports has been validated explicitly; the android path should too.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Lqc6gWzBBnoqvdaQ9HPoEa

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 771885bdee

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".