Meet 4.11.0 shipped an iOS-only SDK change; on iOS bad-call rate dropped 59.4%→31.9% while call success fell 78.7%→73.1%. Android (no SDK change) is flat across the same window — clean control.

v4 bad-call judgment on kVoipSummaryStat: hybrid engine + deliberate-action exclusion + two-layer dedup, inheriting v2's A/V/C+R skeleton. All thresholds TBD.

基于 kVoipSummaryStat 的 v4 bad-call 判定：混合引擎 + 主动行为排除 + 两层去重，继承 v2 的 A/V/C+R 骨架(中文版)。全部阈值待标定。

Per-channel success funnel, failure reasons, and drop-off analysis — table layer plus the instrumentation TODOs needed to build it.

分渠道成功漏斗、失败原因、流失原因 — 表层设计 + 需要的埋点整改(中文版)。

Daily trend of inviteDelay (callee, is_caller=0) and inviteToRing (caller, is_caller=1) over 2026-05-05 to 06-03. inviteToRing avg rose 2.3 → 38.8 s with a single-day spike of 87.3 s on 05/27; inviteDelay p95 stepped up to ~3,050–3,450 s across 05/27–05/31 (peak 3,450.7 s on 05/27). p50 flat throughout.

AE→AE corridor drilldown of inviteDelay (callee) and inviteToRing (caller) over 2026-05-05 to 06-03. inviteToRing raw avg rose 1.72 → 23.09 s while the <36000 s filtered avg moved only 1.72 → 2.74 s (rise is the ≥36000 s tail). inviteDelay filtered avg and p95 both peaked on 05/27 (438.7 s / 1,892.6 s). p50 flat throughout.