fix(agent): transcribe queued voice follow-ups

[bogdanovich]

Summary

• transcribe queued voice follow-up messages before they are converted into steering messages
• preserve normal voice-note behavior when a second inbound message arrives during an already active turn
• add a regression test covering queued voice follow-ups

Problem

When a user sent a second Telegram voice note while the first turn was still running, the follow-up did not go through the normal inbound processMessage() path.

The active-turn branch in AgentLoop.Run() converted the inbound message directly into a steering providers.Message and enqueued it. That bypassed transcribeAudioInMessage(), so queued voice follow-ups could reach the model as raw [voice] media annotations instead of [voice: ...] text.

In practice this made follow-up additions easy to lose. A common example was:

• first voice: what I ate
• second voice while the first turn is still running: and also two pieces of bread

The second voice note could be ignored because it was never transcribed before steering injection.

Fix

Before enqueueing a follow-up into the steering queue, run the same audio transcription step used for normal inbound messages:

• transcribe queued audio media refs
• replace [voice] with [voice: ...]
• preserve transcription placeholder behavior for queued follow-ups too

This keeps queued voice notes behaviorally consistent with ordinary turn starts.

Validation

• go test ./pkg/agent -run 'TestAgentLoop_Run_(AutoContinuesLateSteeringMessage|QueuedVoiceMessageIsTranscribedBeforeSteering)' -count=1
• go test ./pkg/agent -run 'TestProcessMessage_MessageTool|TestNewAgentLoop' -count=1

[afjcjsbx]

this is an annoying picoclaw bug! thanks for the PR, LGTM!