[s2015-turtle]

Problem

Rolling CI for #42 exposed an existing hang in CanTransport.StopsWhileBatchesAreBeingProduced: stop.wait_for(5s) timed out, followed by the overall test timeout.

stop() changed running_ and notified the condition variable without pending_mutex_. The worker could evaluate the predicate as false, miss the shutdown notification before entering its wait, and then sleep indefinitely while stop joined it.

Fix

• Change the shutdown predicate while holding its condition-variable mutex, then notify after releasing the lock.
• Add repeated shutdown coverage around the transition from the last publication into the idle wait.
• Preserve transaction draining and active-frame ordering. No detached threads or timeout relaxation.

Validation

• Actual production transport source compiled locally with temporary ROS type stubs for frame-sink-only tests: all 19 tests passed.
• Original source plus the new test reproduced the hang at stress repetition 180, inner iteration 46 (external timeout exit 124).
• Fixed source passed the identical 200-repetition stress run: 40,000 idle shutdowns plus concurrent-producer shutdown coverage.
• Warning checks and git diff --check passed. All five ROS distributions and the unchanged vcan integration passed on final head5011586269b4be16f5872a5186230cddff4add5b before merge.

This is distinct from #31's permanently blocked publisher case and does not claim bounded shutdown for a sink that never returns.