Embedded shrinker rules from META-INF are selected incorrectly (both with and without r8_extract_embedded_proguard_specs), silently dropping keep rules

[Dolfik1]

Summary

Libraries ship embedded shrinker rules in two locations: legacy META-INF/proguard/ and versioned META-INF/com.android.tools/r8[-from-X][-upto-Y]/. AGP selects between them per jar, with version gating. rules_android gets this wrong in both flag states, and the failure mode is silent loss of keep rules, which under R8 full mode turns into dead-code elimination of live classes.

We hit this migrating a large app from Gradle: R8 removed our Gson-deserialized model classes and the whole push notification pipeline stopped working, with no build error.

Environment: rules_android 0.7.3, Bazel 9, R8 9.1.31.

Case 1: r8_extract_embedded_proguard_specs=false (default)

R8 is invoked with a single merged deploy jar as its program input (rules/android_binary/r8.bzl:128). Modern R8 (9.x) extracts embedded rules from program inputs itself, per archive: if an archive contains any META-INF/com.android.tools/ entries, that archive's legacy META-INF proguard/ rules are ignored (see EmbeddedRulesExtractor in R8).

The merged deploy jar is one archive for all dependencies. kotlinx-coroutines contributes META-INF/com.android.tools/r8*/, which suppresses the legacy rules of unrelated libraries that only ship legacy rules — in our case gson.pro, protobuf.pro, okio.pro, androidx annotations.pro. Merging destroys the per-archive provenance R8's selection relies on.

Repro sketch: deps on kotlinx-coroutines-core + gson, R8 full mode, a class only instantiated via Gson.fromJson(json, Foo.class). unzip -l on the deploy jar shows both META-INF/com.android.tools/r8-from-1.6.0/coroutines.pro and META-INF/proguard/gson.pro; the resulting dex has the coroutines ServiceLoader optimization applied (proving R8 did extract from the merged jar) but Foo is removed.

Case 2: r8_extract_embedded_proguard_specs=true

tools/android/proguard_extractor_lib.py concatenates everything under both META-INF/proguard/ and META-INF/com.android.tools/ with no version gating and no precedence. Concrete consequences with kotlinx-coroutines at R8 9.x:

• META-INF/com.android.tools/r8-upto-3.0.0/coroutines.pro is applied even though the current R8 is 9.x. That file contains -keep class kotlinx.coroutines.android.AndroidDispatcherFactory {*;}, which defeats R8's ServiceLoader optimization — the exact thing the versioned directory scheme exists to prevent (R8 has built-in support for ServiceLoader, Proguard rules need update Kotlin/kotlinx.coroutines#3111).
• Legacy META-INF/proguard/coroutines.pro is applied on top, plus META-INF/com.android.tools/proguard/ files, which AGP never applies.

The existing unit tests (testR8RulesVersionedSubdirs, testLegacyAndR8RulesCombined) assert this concatenation behavior, so the bug is currently encoded as expected behavior.

Reference semantics

AGP: ExtractProGuardRulesTransform + FilterShrinkerRulesTransform (per-jar; tools dir existence suppresses legacy; only r8* subdirs, from inclusive / -upto- exclusive against the R8 version; union of all matching dirs; com.android.tools/proguard/ ignored). R8 side: EmbeddedRulesExtractor implements the same shape per archive. AAR handling differs too: AGP reads META-INF/com.android.tools/** from classes.jar and falls back to the AAR's proguard.txt only if none found; rules_android's AAR extractor concatenates both.

What we did

We're running a local patch (Dolfik1@c0e1a1a) that:

1. Rewrites the extractor to per-jar AGP semantics and feeds it the individual library runtime jars (via a params file) instead of the merged deploy jar, with --r8_version for gating (plumbed as a toolchain attr). It also emits a manifest listing every discovered rule entry with a verdict (SELECTED / SUPPRESSED_BY_TOOLS_DIR / VERSION_EXCLUDED / NON_R8_DIR), which turned out to be very useful for verifying parity with the Gradle build.
2. When the flag is on, strips META-INF/proguard/** and META-INF/com.android.tools/** from a copy of the deploy jar before it is passed to R8, so R8's own extractor can't re-apply a wrong merged-archive selection on top. The original deploy jar is untouched for its other consumers.
3. Fixes the AAR extractor to match AGP (classes.jar tools rules first, proguard.txt only as fallback).

Would a PR along these lines be welcome? If yes: one PR or split (extractor+gating / strip step / AAR path)? And do you want the new behavior gated behind anything beyond the existing r8_extract_embedded_proguard_specs flag?