fix(menubar): a take on Today's mic shows on the Models section's Dictation row

[spokvulcan]

Summary

The menu's Models section reads the Dictation row's Listening and Transcribing from the container's Voice Inputs. Its hand-written list, written in #604, missed Today's, added in #624.

 a take on Today's mic        Dictation row
-  holding the mic            Idle
+  holding the mic            Listening
   Whisper decoding           Transcribing     (read from the shared engine, already right)
-  the Proofread Pass         Idle
+  the Proofread Pass         Transcribing

Every Voice Input in the container now comes from one factory, which registers it:

 DependencyContainer
+  makeVoiceInput()     a Voice Input on the shared mic, Whisper and Proofread Pass, added to voiceInputs
   agentVoiceInput    = makeVoiceInput()
   captureVoiceInput  = makeVoiceInput()
   panelVoiceInput    = makeVoiceInput()
   todayVoiceInput    = makeVoiceInput()

 modelActivity(), the Dictation row
-  [agentVoiceInput, captureVoiceInput, panelVoiceInput]
+  voiceInputs

The four Voice Inputs take the same arguments as before. modelActivity() no longer creates the lazy ones it reads on the menu's first open: it reads only those made.

Not in this PR: the onboarding's Dictation chapter builds its own Voice Input in its view, outside the container. It has no Proofread Pass, and its Whisper decode shows through the shared engine.

Evidence

• Before: Today's Voice Input built directly, as on main. Both new tests in MenuModelsSectionTests fail:

  FAILED theContainerBuildsEveryPushToTalkInOnePlace()
      count of "AgentVoiceInputController(" in DependencyContainer*.swift → 2, expected 1
  FAILED theDictationRowHearsEveryPushToTalk()
      container.voiceInputs.contains { $0 === container.todayVoiceInput } → false
  

  After: MenuModelsSectionTests and LoadedModelsTests: 8 passed. Whole target: 3,631 passed, 0 failed, 7 skipped. scripts/check.sh passes.
• Not checked in the running app: the mic is a mouse hold, so the menu can't be open while it records.

Merge Danger

Door: two-way

Wiring only: no stored data, settings or API. A revert puts the list back.

Blast Radius: menu

The Models section's Dictation row. The four Voice Inputs are built with the same arguments, now in one place.

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.