v0.3.4 — notifier test no longer flakes under load

Test-only release, shipped so the fix reaches the AUR package.

test_audio_spawned_before_blocking_notify_send failed roughly 1 run in 45 on a loaded machine. The product code was never wrong — audio has always been spawned before the blocking notify-send. The test was: it compared the order two shims appended to a log, but mpv is spawned detached, so under CPU load the kernel can schedule it after the foreground notify-send has already written. The log inverts while the spawn order stays correct.

This is worth a release because the suite runs in the AUR PKGBUILD's check(), on the machine of everyone who builds the package — and a build machine is loaded by definition, so the flake surfaced as occasional install failures rather than as a red CI run someone would notice.

The test now asserts the property directly: with notify-send blocked for 3s, the adhan must be audible within 2s, while it is still blocking. Verified by mutation — moving the audio call after notify-send makes it fail — and by 60 runs under 4-way CPU load with zero failures.

No behaviour change for users.