[tara-bogavelli]

Adds opt-in interruptions, backchannels, and speculative interruption generation to the cascade simulator in #214. This draft targets feat/cascade-user-simulator; its diff contains only the experimental layer so the base can merge while behavioral testing continues.

Includes goal-aware listener checks, transcript-change gating, one interruption per assistant turn, relevance checks for pre-generated candidates, shared phrase-audio caching and localization, and listener/interruption diagnostics. Tick-driven barge-ins discard unheard buffered audio and request provider truncation at the played position. All behavior flags default off; self-correction remains removed.

Validation:

• pytest tests -q: 2,287 passed, 52 skipped, 3 xfailed, including the inherited full-loop tests with behavior flags off.
• pre-commit run --all-files: passed.
• With all flags enabled, bounded Pipecat debug run produced three ordinary caller turns and one interruption; OpenAI Realtime produced two ordinary caller turns, one interruption, and one backchannel.
• Both runs reached the 90-second conversation limit. Final speech-fidelity validation failed because the configured Vertex credentials could not be loaded. These smoke runs exercise the paths but do not establish end-to-end behavioral quality.
• Compared the reconstructed stack against the saved original implementation integrated with current main: production Python behavior is preserved, apart from the intentional simulation-version change. Added documentation and shared regression coverage are retained.

After #214 merges, rebase only this PR's commit onto main and retarget this PR, especially if the base is squash-merged.