the runner decodes each finished test's stdout and stderr strictly. nodes inherit the test process's stderr, so a node that aborts can put raw bytes there, and the decode raises UnicodeDecodeError inside TestHandler.get_next(). that takes the whole run down before it reports which test produced them, so what you see is a traceback in the runner and the real cause is invisible.
moves the decode into read_captured_output() and decodes with errors='replace'. the pass criterion is untouched: a test with anything on stderr still fails, it just gets named now.
reproduced on master with a scratch test writing b"Assertion failed\x80\x81 at 0x7f\n" to stderr and exiting 1:
before
UnicodeDecodeError: 'utf-8' codec can't decode byte 0x80 in position 16: invalid start byte
(no results table at all)
after
feature_zzzbadbyte.py | ✖ Failed | 0 s
Assertion failed<replacement chars> at 0x7f
two unit tests cover the helper. feature_framework_unit_tests.py previously loaded only test_framework modules, so it now also loads test_runner by name, 33 tests instead of 31. restoring the strict decode fails test_captured_output_survives_undecodable_bytes with the same UnicodeDecodeError the runner used to die on.
found while chasing a real failure on a debug build of a downstream fork, where a node died and the run reported a decode error instead of naming the test. this does not prevent that failure, it only makes it diagnosable in one step instead of several.
litecoin carries the same line, and dogecoin's older copy of this runner does too.