[tara-bogavelli]

Adds a standalone cascade user simulator that combines streaming STT, a caller LLM, and Cartesia TTS. The caller uses the shared persona/goal prompts and end-call tool, waits for replies, and records transcripts, audio boundaries, clean caller audio, and per-tick diagnostics.

Includes both real-time streaming and tick-driven OpenAI Realtime adapters, pacing selection, background-noise/connection perturbations, transcript fallback, and the latest inactivity/provider-stall fixes. Also removes unused Backend/Role scaffolding and documents simulator setup.

This PR is the independently mergeable base of a two-PR stack. All listener decisions, interruptions, backchannels, speculative generation, localized phrase caching, and barge-in truncation are in dependent draft PR #228, which targets this branch. Those flags and execution paths are absent here.

Validation:

• pytest tests -q: 2,192 passed, 52 skipped, 3 xfailed.
• pre-commit run --all-files: passed.
• Added full-loop regressions for reply waiting, partial transcripts, hang-up, contiguous inactivity, and provider-stall artifact preservation.
• Bounded Pipecat debug run exchanged four caller turns before the 90-second conversation limit; the smoke configuration omitted mandatory judge deployments, so final validation did not pass.
• OpenAI Realtime base smoke failed during the upstream connection handshake. The dependent branch subsequently exercised this adapter live, including out-of-turn speech, but does not substitute for successful base end-to-end validation.

Rebuilt on current main (4cd0028b). The original branch tip is preserved locally as backup/cascade-user-simulator-20260916 (8467d0a4).

[JosephMarinier]

Closing in favor of #240