[twstagg]

Symptom

Once configure succeeds (see #5431), the utils suite is what fails in a loaded build:

test/utils/sleeper_thread.cpp:79: FAILED:
  REQUIRE( run_in_subprocess(run_control_flag_stress) == 0 )
with expansion:
  -1 == 0

assertions: 756 | 755 passed | 1 failed
3/3 waybar:utils    FAIL    5.52s   exit status 42

It looks like a SleeperThread defect. It is not: it is the test's own watchdog, and it only fires when the machine is busy, so it reads as a flake.

Cause

run_in_subprocess starts the task under alarm(5) and leaves SIGALRM at its default disposition, which terminates the child. The parent then sees !WIFEXITED(status) and returns -1 — the same value it returns for a fork failure or a crash. So a child that simply needed more than five seconds is reported identically to a genuine logic failure, and the message gives no hint that a timeout happened at all.

The stress loop is wall-clock hungry rather than slow: it runs 200 iterations of thread creation, 100 wake_up() calls, stop() and join(), so under scheduling pressure every iteration pays for context switches. Measured on a 16-core machine saturated with 16 busy loops, the loop still completed, taking up to 30.8 s:

tree	runs under load	failures	slowest run
master	40	19	5063 ms
master + #5431	40	25	5059 ms
this change	40	0	30775 ms

Every failure sits at ~5.0 s, i.e. exactly at the alarm. The five-second budget is therefore an undeclared performance assertion attached to a correctness test, and it fails on shared CI runners and on any machine doing parallel builds — the Nix build in particular.

Change

• Give the watchdog a generous budget (60 s, comfortably above the 30.8 s worst case measured under full-core saturation) so it guards against a hang, which is its actual purpose, instead of also asserting speed.
• Install a handler that exits with a dedicated code (124) on expiry, so an expiry is distinguishable from a crash instead of arriving as "killed by a signal".
• Report the outcome in the failure message, so the next person sees watchdog expired: the stress loop did not finish… rather than -1 == 0.

The stress magnitude is deliberately unchanged — the volume of concurrent wake and stop calls is the point of the test.

Verification

• utils suite: 772 assertions in 44 test cases, all passing.
• 40 runs of [sleeper_thread] under 16-core saturation: 0 failures (was 19–25 of 40).
• Full Nix build with tests enabled: Ok: 3, Fail: 0.

Depends on #5431 only in the sense that it is what you hit next; the two changes are independent.