Tap the hotkey to toggle dictation

[TerrifiedBug]

dictation.tap_to_toggle makes a tap start recording and a second tap stop it, instead of holding the key down. A press held past half a second still ends on release, so push-to-talk keeps working and the mode never latches the mic open on someone's muscle memory. Config-only and read at the point of use, like the other dictation niceties, so it hot-reloads; the menu bar's idle line and the boot banner say which verb applies.

It also fixes a state race the new mode makes easy to hit. A press landing while the previous press's transcription was in flight had its indicator, overlay and capture torn down by that older press's completion — the pill vanished and, because dictating went false under it, the new capture ran on with no release to stop it. Presses now carry a generation and finishDictation only cleans up for its own.

Verification

Run in the foreground against synthesized fn flagsChanged events (real event tap, real mic, real model):

scenario	result
default config, 1.04 s hold	● dictating / ○ captured 0.79s / → 63ms — baseline unchanged
default config, 186 ms tap	○ captured 0.09s — a tap does not latch in hold mode
tap_to_toggle saved while running	config reloaded, menu line reads idle · tap fn to dictate
toggle mode, tap … 1.7 s … tap	one ● dictating, one ○ captured 1.80s spanning both taps, pill up and state ● listening with the key released
toggle mode, 975 ms hold	○ captured 0.90s on release — push-to-talk preserved
race: third press 36 ms into the second press's transcription	● dictating logged before the older press's → 71ms, pill still on screen, and the next tap's ○ captured 0.79s matches only the third press's window
tap_to_toggle set back to false mid-latch	still ● listening; one press stops it (○ captured 1.88s)

Numbers

yap bench --audio on LibriSpeech clips, M4, p50 of 7, parakeet-tdt-ctc-110m, before/after:

clip	before	after
2 s	34 ms	36 ms
5 s	43 ms	40 ms
10 s	49 ms	51 ms
20 s	75 ms	76 ms

Noise in both directions, as expected from a change that does not touch the transcription path. The key-down path gains exactly one config read — measured at 25.6 µs p50, 71.8 µs p99 against a 60 ms engine.start() — and the release path gains a stored-bool check and one Date() comparison.