v0.2.0

• Package 0.2.0. Offline voice extension: energy VAD on synthetic PCM, turn-taking state machine, barge-in cancel, SIMULATED 100-call latency report (p50/p95 labelled SIMULATED, with a seeded bootstrap interval), webhook v2 (signed timestamp, retry classes, full jitter, tool deadline, W01–W12), correlation logs, and trace.
• python -m voice_cascade eval-voice writes docs/voice/SIMULATED_LATENCY.md. live --dry-run prints an [EST] cost and sends no request. Silero is an optional marker, off by default, and does not download weights.
• Decisions D1–D5 stay open. The offline defaults are recorded in docs/voice/OPEN_DECISIONS.md (L1, DSP audio, bundle D, manual Silero, v1 signatures accepted and deprecated).
• The 49-script SIMULATED figures are unchanged: p50 896 ms and p95 1004.5 ms, from package 0.1.9. They are not compared with the new cohort.
• Design note for VAD/barge-in, per-stage latency and webhook failures (the note itself is design; the offline code is the 0.2.0 bullets above).
• README opens with a three-line TL;DR under the title and badges.
• Docs only. The SIMULATED eval figures stay p50 896 ms and p95 1004.5 ms. The cited package version is 0.1.9 and the cited git revision is 4fb26ef5a281cc7b325c75c5f6e0f7a96e9a16fa on main.
• One measured xAI run (2026-10-17/18, repo HEAD 84cfb1a, n=20, not an SLA) is in the README table "Measured with real providers", with evidence in docs/live/ and notes in docs/LIVE_RUN.md. The scratch adapter is not in the repo. A short English demo and captions are in docs/demo/.
• The availability pattern list and the four handwritten recall sets (11/20, 15/20, 14/20, 16/20) are in docs/HEURISTIC_RECALL.md. The intervals overlap; those rates are not an improvement. Recall after the latest patterns is not measured yet.
• Vulnerability reports use GitHub private vulnerability reporting. No email address is listed.