[bmc08gt]

Unit tests takes about 16 minutes, and on warm code/cash pushes most of that is work the build cache should have skipped. Run 41547433541 restored the previous commit's cache and still executed 2,821 of 4,584 actionable tasks. Three causes:

1. KMP modules built release copies of their Android dependencies. The single-variant KMP Android target picks one build type from each Android-library dependency, and AGP defaults that to release. flipcashTestDebug therefore compiled 25 Android libraries twice, debug for their own tests and release for the KMP host tests: 444 release-variant tasks, compileReleaseKotlin included. Both KMP convention plugins now set localDependencySelection { selectBuildTypeFrom = [debug, release] }. The task graph drops from 5,932 tasks to 5,426, with no release tasks left.

This changes what the KMP host tests run against: debug builds of those libraries instead of release builds. The debug app already runs on the debug builds. The app's own releaseRuntimeClasspath resolves by the app's attributes and is untouched.

2. The app read the commit SHA while configuring. bugsnagBuildId ran git rev-parse HEAD at configuration time and went into every variant's manifest. Each commit changed the debug manifest, re-ran :apps:flipcash:app:testDebugUnitTest, and invalidated the configuration cache. Bugsnag is off in debug, so debug now gets the constant "debug". The SHA is a lazy provider wired only into non-debug variants and the mapping upload.

versionCode (git rev-list --count) has the same shape, but CI's fetch-depth: 1 checkout always counts 1, so it doesn't vary there and this PR leaves it alone.

3. The configuration cache never survived a run. Every run logs "no cached configuration is available" and spends about a minute configuring (23:43:08 to the first project task at 23:44:05 in that run). setup-gradle saves the configuration cache only with cache-encryption-key, since the entry holds the provisioned API keys. Both jobs now pass secrets.GRADLE_ENCRYPTION_KEY. Until that secret exists, the input is empty and setup-gradle skips the configuration cache as it does today.

Measured locally, fresh checkout per run against a warm local build cache, generateEmojiList flipcashTestDebug:

	Executed	From cache	Compile Kotlin executed	Unit-test tasks executed
Before, empty commit	2,121	2,468	2	1
After, empty commit, versionCode 1 as on CI	1,948	2,195	0	0
Before, one string added to apps/flipcash/core	2,879	1,766	102	70
After, same string	2,705	1,496	96	70

The remaining executed tasks on an empty commit are AGP's non-cacheable packaging and copy tasks.

The last two rows are what CI saw on run 41547433541: its counts match the "before, one string" row (102 compile, 70 unit-test tasks executed). That commit added strings to apps/flipcash/core, which most modules depend on, so their R classes change and their compile and test tasks re-run legitimately. This PR doesn't change that; moving resources out of core or narrowing what depends on it is a separate change.