Fix #12912: use a dumb fallback terminal for all build-thread reentrant calls - #12925

Summary

Fixes Maven 4.0.0-rc-6 hanging on startup with JDK 25 in Linux container environments, and fixes a flaky FastTerminalReentrancyTest.usingTheTerminalFromTheConsumerDoesNotDeadlock deadlock observed across JDK 17/21/25 on ubuntu CI runners (example, example, example).

Root cause: FastTerminal only guarded writer() and getType() against build-thread reentrancy. JLine 4.4.0's FFM provider initialization (CLibrary.<clinit>) and AnsiConsole.systemInstall() can reach back through other terminal methods (getSize, getName, encoding, etc.) on the build thread, causing a deadlock where the build thread waits on the CompletableFuture it is itself computing.

Fix: Move the isBuildThreadWaitingOnItself() guard into getTerminal() — the single choke point all delegate methods pass through — backed by a pre-built DumbTerminal. This covers all methods at once, including any new ones JLine may add.

Improvements over #12921:

• The DumbTerminal's output stream is wrapped with FilterOutputStream so that close() flushes instead of closing the captured System.err — DumbTerminal.doClose() cascades to its writer's underlying stream, and closing System.err would silently break stderr for the process
• Removed the now-redundant per-method guards on writer() and getType(), along with the unused fallbackWriter field and fallbackWriter() method — the single guard in getTerminal() makes them dead code

Test plan

• Existing tests for writer() and getType() still pass (4 tests, 10/10 stress runs)
• New test arbitraryTerminalMethodsFromTheBuilderDoNotDeadlock verifies getSize() from the builder callback
• New test arbitraryTerminalMethodsFromTheConsumerDoNotDeadlock verifies getName() from the consumer callback
• CI passes on all platforms

Supersedes #12921.

🤖 Generated with Claude Code