Fixed
- Tone detection on paging audio (tone-then-voice) — Shared decode path merges Discover and production Detect so quiet lead-in tones are not lost when voice follows; 20 s FFT window and 3 s peak reference for silence gating; auto-learn keeps looser Discover thresholds.
- Pending tone stacking — Merged pending tones use the most recent tone page as the primary attach label.
- Tone auto-learn pairing — Default A-tone 0.5–1.2 s, B-tone 1.5–4.0 s; earliest valid A with longest following B per call.
Added
- Analyze tone history diagnostics — Per-call discovery details and partial patterns in the API for admin review.
- Regression tests — Production export tests for 78 FD DISP and LFD sample audio (Discover + learned A/B detect).