test: 100% line and branch coverage of the whole library

[tony19]

Summary

This extends the coverage gate from #477, which covered only the Kotlin Android layer, to the whole library. CI now fails unless the unit tests cover every line and every branch of every class, Java and Kotlin. It also adds the tests that meet the gate.

	Before	After
Lines covered	9,060 / 11,918 (76%)	11,881 / 11,881 (100%)
Branches covered	3,105 / 4,461 (70%)	4,426 / 4,426 (100%)
Files with gaps	296	0

These numbers hold for both flavors (jdk11Debug, jdk8Debug) on both CI JDKs (17 and 21).

Build (build:)

• Gate: coverageReport<Variant>, coverageVerification<Variant> and verifyCoverage replace the androidLayer* tasks. They cover ch.qos.logback.** and org.slf4j.impl.** with the same rule as before: 100% of each class's lines and branches.
• scripts/coverage-gaps.py: lists what a JaCoCo XML report leaves uncovered, one file per line.
• CI order: CI writes both variants' reports first, prints their gaps to the log and the job summary, and then runs the gate. A failing gate therefore says exactly what is missing.
• Docs: .github/copilot-instructions.md and the phase-start skill now mention verifyCoverage.

Tests (test:)

• Scale: about 290 test files, with +37k lines of new or extended JUnit 4 tests.
• What they assert: each test asserts observable behavior: return values, state, output, status messages and exceptions. Every package was mutation-checked by an independent review pass: covered code was changed, and at least one test had to fail. Surviving mutants got new tests.
• Deterministic coverage: comparing CI runs showed lines covered in some runs and not in others (timing), in SocketNode, SocketReceiver, AsyncAppenderBase, DefaultSocketConnector and TimeBasedArchiveRemover. They now have deterministic tests, so the gate can't flake.
• Off Robolectric where possible: several tests that used Robolectric only for the default context now run on the plain JVM. SocketAppenderMessageLossTest timed out once on CI inside its 3 s budget. It now runs on the plain JVM with a 10 s hang guard.
• Compiles against Android's Java library: CI compiles unit tests against Android's Java library, not the full JDK. For example, SMTPAppenderBaseTest reads the part's content type without naming javax.activation.DataHandler, which implements java.awt.datatransfer.Transferable.
• JDK dependency to watch: LoaderTest's permission tests rely on AccessController working, which it does on CI's JDK 17 and 21. A comment next to them notes that JDK 24+ will need attention.

Bugs found and fixed (fix:)

Each fix has a test that fails without it.

• CyclicBuffer.resize(): it copied nulls, or threw ArrayIndexOutOfBoundsException, on a buffer that get() had partly drained. Resizing to the current element count was also ignored. The same bug is in upstream.
• OptionTokenizer: a pattern whose quoted option ends in an escape, such as %x{'a\', threw StringIndexOutOfBoundsException out of PatternLayoutBase.start(). It now throws ScanException, which becomes an error status.
• TimeBasedArchiveRemover.findEmptyDirs(): when cleanup deleted a directory's only child, it left the now-empty parent until a later pass. It now removes it in the same pass.
• SimpleSocketServer.doMain(): it ignored its serverClass argument, so SimpleSSLSocketServer.main started a server without SSL. The divergence from upstream is commented at the site.

Unreachable code removed (refactor:)

Where a branch could not be reached by any input, subclass, configuration or mock, it was removed or folded, without changing behavior or public/protected API:

• Enum switches: implicit default: branches of switches that cover every constant were folded into the last case: Compressor, FixedWindowRollingPolicy, TokenStream, the subst Node/Tokenizer/NodeToStringTransformer, ActionUtil, NestedBasicPropertyIA/NestedComplexPropertyIA, PropertySetter. This form covers fully under both javac 17 and javac 21, which compile enum switches differently.
• Null checks that can't fail: ElementPath, FormatInfo, ReconfigureOnChangeTask, PropertySetter, Introspector.
• Catches for exceptions that can't be thrown:
  • OptionHelper: SecurityException. Android has no SecurityManager, and JDK 18+ won't install one at runtime.
  • HardenedObjectInputStream: merged its duplicate reflection catches.
  • StringToObjectConverter: merged two identical catch blocks.
  • AbstractIncludeAction: a MalformedURLException catch that only printed a stack trace. The exception is now handled together with JoranException in begin().
  • SaxEventRecorder: the "can never be reached" throw.
• ThrowableProxy: calls Throwable.getSuppressed() directly instead of through reflection (API 19+, as upstream 1.3 does).
• ContextInitializer: dropped its always-true search flags.

Test seams (package-private, no public API change)

• AbstractSocketAppender: the SSL handshake-failure delay is a field with a package-private setter, so tests don't sleep 30 s. The dispatch loop moved into its caller so that JaCoCo can mark it covered.
• TimeBasedFileNamingAndTriggeringPolicyBase.newActiveFile(): lets a test supply an unreadable active file. Before, the only test for that branch withdrew read permission, which does nothing for root or on Windows, so the gate would have depended on who ran it.

Tried and reverted

• Prudent mode: routing OutputStreamAppender.subAppend() through writeOut(), as upstream 1.3.0 does, would make prudent mode actually lock the file. The final review showed this drops events when two threads log at once: FileAppender.safeWrite() takes the FileChannel lock without holding the appender's lock, so the second thread gets OverlappingFileLockException. It also lets an interrupt close the channel for good. Upstream fixed this later by locking the stream around the file lock. That behavior change belongs in its own PR, so the 1.2 behavior stays, and writeOut()/safeWrite() are covered by calling them directly. FileAppenderTest.prudentAppenderKeepsEveryEventLoggedConcurrently pins the restored behavior: with the reverted change it loses half of the events, every run.

Architecture: no package, dependency or published-output changes; docs/architecture/ is untouched.

Deferred work considered:

• Open issues: When Device Restart happens, some logs are not saved to file using RollingFileAppender #371, can not write to the file that under the android getExternalFilesDir()  #366, Excepton in kotlin app with jetpack compose, only in release mode #352, Can't Create Logger: DateTokenConverter not found? #344, EXT_DIR value depends on Android version #223 and SizeBasedTriggeringPolicy fails for relative file paths #135 are adjacent (M1.1–M1.3), not in scope. This PR only adds tests around those classes, so they stay open.
• Prudent-mode locking: see Tried and reverted above; it's left for a follow-up.

🤖 Generated with Claude Code

https://claude.ai/code/session_0157UjTSjJyouKBfY7Zwxeig

[tony19]

Build (JDK 17) and Build (JDK 21) fail at the new Coverage gate step, which is expected at this stage. Every earlier step passes: assemble, bytecode check, lint and both flavors' unit tests.

This PR adds the gate, and the tests that satisfy it haven't landed yet. The Coverage gaps step lists what is uncovered:

• Both flavors: 296 files, with 9,060/11,918 lines and 3,105/4,461 branches covered.
• Timing-dependent lines: comparing the four runs (JDK 17/21 × jdk11/jdk8) showed lines in SocketNode, SocketReceiver, AsyncAppenderBase, DefaultSocketConnector and TimeBasedArchiveRemover that are covered in some runs but not others. These get deterministic tests so the gate can't flake.

The tests are being written package by package and will be pushed to this branch. CI should go green once they cover every line and branch.

---

Generated by Claude Code