[nirwester]

GCC 15 emits a (I think) false positive -Wnull-dereference when copying the subscriber vector in ThreadSafeSignal::notify(), breaking downstream builds that enable -Werror, including Nav2 (an example is here).

Copy-constructing the snapshot before moving it into the local vector should avoid the problematic assignment path. Locking, callback behavior, and API/ABI remain unchanged.

Validated with GCC 15.2 at -O2 and -O3, on existing callback and blackboard tests, and ASan/UBSan checks.
If I get a bit of time I can add a report for GCC too.

[facontidavide]

CI is unhappy

[nirwester]

Seem race conditions unrelated from the change. According to Claudio (I didn't have time to try to follow the flow myself):

1. The test signals startup too early. starts++ happens inside onStart(), before executeTick() publishes RUNNING. The main thread can halt the still-IDLE node, then the in-flight tick overwrites that halt with RUNNING. Forcing this ordering reproduces the failure consistently.
2. The worker can restart an already halted tree. A diagnostic run captured:
   before_halt=RUNNING after_halt=IDLE at_timeout=RUNNING starts=2
   In tickRoot(), the halt check precedes the wake-up poll. A halt arriving between those operations can be consumed as an ordinary wake-up, causing another tick and restarting the root.
3. The condition variable can miss a notification. WakeUpSignal::emitSignal() updates its predicate without holding the waiter’s mutex. Notification can arrive between the predicate check and entering the wait. I reproduced the waiter sleeping until its timeout despite receiving a signal; locking the producer’s predicate update prevented that failure.

I also tried to have him propose patches for these, but they looked quite invasive, and I don't want to PR AI rubbish :)

[SteveMacenski]

BTW I filed a ticket this would resolve: #1288