Reliability: wire the Temporal voice provider (pipecat_provider) end-to-end

[teetangh]

Context

The Temporal production path (apps/worker/activities.py → resolve_via_voice → get_voice_provider()) selects PipecatProvider (apps/voice/pipecat_provider.py) per settings.yaml (voice.provider: pipecat) — but that provider is a non-functional scaffold:

• _LiveTranscriptCapture.process_frame is a ... no-op (lines 148-152) — transcripts stay empty, breaking the cross-modal HandoffContext claim.
• _classify_outcome relies on the LLM literally saying "deal agreed" in text (the streaming module long since moved to end_call tool calls).
• recording_url and cost_usd are hardcoded placeholders (cost_usd=0.0 at line 102).
• It references a TwilioTransport import that doesn't match the shipped Pipecat 1.x API.

Meanwhile the real, working pipeline lives in apps/voice/streaming.py behind the FastAPI WS endpoint.

Implementation plan

1. Reuse, don't duplicate: refactor pipecat_provider.py to drive the existing HTTP flow — call POST /voice/streaming/dial-init, place the call, then await the session result — rather than constructing a second Pipecat pipeline. Extract shared session-result plumbing (outcome, transcript, cost) into apps/voice/session_result.py.
2. Transcript capture: in the streaming pipeline, subscribe to TranscriptProcessor (or aggregate TranscriptionFrame / assistant TextFrames via the existing context aggregator) and persist turn-by-turn transcript on the session record.
3. Outcome from tools: session outcome comes from the end_call(outcome=...) tool args (handle_end_call, streaming.py:537-548) — expose it on the session result; delete _classify_outcome text matching.
4. Recording URL: capture from Twilio's recording status callback (call placed with record=True) and attach to the session result.
5. Cost: take total_cost_usd from the P0 VoiceCallMetricsObserver summary instead of 0.0.
6. Borrower phone lookup: replace BORROWER_PHONE__{id} env vars (pipecat_provider.py:110-124) with a proper lookup (settings/DB), keeping the env override for dev.
7. Tests: activity-level test with the mock provider still green; integration smoke running the Temporal workflow against the simulator-backed streaming endpoint with populated transcript/outcome/cost.

Acceptance criteria

• CollectionsWorkflow completes end-to-end with voice.provider: pipecat against the simulator: non-empty transcript, outcome sourced from the end_call tool, non-zero cost_usd, recording URL populated on real calls.
• No duplicate pipeline-construction code between pipecat_provider.py and streaming.py.

Dependencies

• P0 fixes PR (metrics observer for cost).
• Observability: transcript + offer-event persistence with compliance audit trail #9 transcript + offer-event persistence (shared session-result storage).