[Epic] Telephony, channels, and runtime reliability #82

Parent roadmap: #76

Problem

QuickVoice supports Twilio and Telnyx number purchase/binding, LiveKit SIP, web preview/widget sessions, recordings, outbound calls, and a configurable STT/LLM/TTS runtime. Production users also need existing-carrier connectivity, sophisticated routing and handoffs, channel-consistent context, provider fallback, and transparent behavior during partial outages.

Outcome

Customers can connect or port telephony, define inbound and outbound routing, move conversations safely between AI and humans, operate across phone and web channels, and maintain a controlled degraded experience when a carrier, model, speech provider, tool, or region fails.

Required capabilities

• BYOC SIP trunks, credential/IP/TLS/SRTP options, number porting workflow, caller-ID verification, routing rules, business hours, regional ingress/media, health checks, and failover.
• Warm and cold transfer, queues, ring groups, skills/availability routing, scheduled callbacks, live takeover, supervisor whisper, and concise structured handoff context.
• Configurable provider/model policy using language, region, health, latency, quality, capability, price, privacy, and capacity.
• STT/LLM/TTS fallbacks with compatible voice/config mapping and observable reason codes.
• Web voice/widget and mobile-friendly SDK lifecycle, authenticated sessions, DTMF, voicemail, answering-machine behavior, hold/resume, mute, hangup, and optional SMS follow-up.
• End-to-end concurrency, timeout, cancellation, retry, and capacity controls.

Child issues

• #94 — BYOC SIP, number porting, routing, failover, and regional media controls
• #99 — Warm transfer, queues, callbacks, live takeover, whisper, and handoff summaries
• #102 — Provider/model routing, fallback, multilingual detection, and cost/latency policies
• #106 — Web/mobile voice, widgets, SMS follow-up, DTMF, voicemail, and call controls

Acceptance criteria

• Routing decisions are deterministic, versioned, explainable, and testable without placing a paid call where possible.
• Provider or regional health degradation can shift eligible new sessions without interrupting incompatible in-flight calls silently.
• A handoff communicates consented context and an accurate summary while respecting retention and redaction policy.
• Missing human availability follows configured queue/callback/voicemail/fallback behavior.
• All call-control actions are authorized, idempotent, and auditable.
• Channel SDKs expose documented connection and failure states and recover from transient disconnects within policy.
• SIP, transfer, routing, failure injection, fallback, concurrency, DTMF, voicemail, session auth, and redaction tests are included.

Boundaries

Emergency calling, lawful caller-ID presentation, number eligibility, media residency, and carrier behavior vary by provider and jurisdiction and must be surfaced as explicit configuration and documentation boundaries.