[johnmatthewtennant]

Summary

Add berd-call vad-replay INPUT.json [--candidates CANDIDATES.json] to compare continuous-recognition VAD settings offline using the production SpeechGate. JSONL reports distinguish human-checked labels, inferred labels, unknown capture intervals, onset misses and delays, fragmentation, and needless reply holding while a reply is waiting.

This is the replay foundation for call recording and tuning. It does not change live profiles or simulate audible playback. berd-call vad-replay events CAPTURE_DIR also reads capture manifests/events and separates request-wait overlaps with VAD and pending recognition. Per-frame capture trace conversion and native recognition-policy trials remain incomplete. No architectural laws change.

Related issue

Stacked on independent VAD profiles and timing.

Testing

• Twelve unit/process tests cover threshold comparison, production gate behavior, mode/epoch changes, gaps, label provenance, reply-wait penalties, latency/fragmentation and the actual CLI.
• Full Linux cargo test -p berd-call --no-fail-fast: the same three failures occur on unmodified base 681fea86 and this change (management usage, synthesis usage, unresolved handoff expiry). New replay tests pass.
• Library Clippy passes with warnings denied. All-target Clippy reports existing cross-platform dead-code/lint failures.
• just tauri-check is blocked by missing Linux gobject development metadata on the disposable workstation.
• CLI process example compares a recorded 0.5 threshold with a 0.7 candidate on the same 0.6-confidence speech frame: reported missed speech changes from 0 to 16 ms, input stays byte-identical.

[johnmatthewtennant]

🤖 Consolidated into #426 with capture, device profiles and timing diagnostics. Continued implementation and review belong there.