[trufae]

The two failing Linux CI tests exercise pdgj/pdgx with r2ghidra.timeout. The child flushes directly to stdout, bypassing the parent console where JSON selectors, grep counts and redirections are applied. The tests also set the plugin option with -e before plugin initialization, producing an unknown-variable error on radare2 6.2.4.

Send the child's raw console buffer through the pipe and append it to the parent console after successful completion. Drain the pipe while enforcing a monotonic deadline so output larger than the pipe capacity cannot deadlock. Reset the inherited child buffer, handle interrupted/partial I/O, close descriptors on fork failure, and use _exit to avoid flushing inherited stdio. Set the timeout inside the existing test command sequences after plugin initialization.

Validation against the same radare2 6.2.4 release used by CI:

• Full make -C test: 189 passed, 0 failed.
• Original worker code with corrected test setup: 6 passed, 2 failed; fixed worker: 8 passed.
• C, JSON and XML output matches direct execution byte-for-byte (2,045 / 55,293 / 95,839 bytes).
• Checked buffered output without duplication, JSON file redirection without stdout leakage, and a forced 1ms timeout followed by successful decompilation.

The macOS ARM failure in the master run was a transient Could not resolve host: github.com while cloning ghidra-native. That job has been retried; no platform build workaround is included.