Research scope

Captures the Duet/Focusrite stream-stability investigation so contributors can review the cursor, replay, and recovery assumptions together.

What changed

• Adds CoreAudio write-frontier → TX-preparation signalling and telemetry.
• Adds compact MCP-ring fault records, including the precise RxSequenceReplayReader::TryRead() rejection reason.
• Adds a TX data-horizon burst simulator and updates buffer-geometry verification.
• Documents the observed cursor invariants, replay-horizon failure, FW-64 callback overwrite, and unsafe recovery escalation.

Current conclusion

This is intentionally research, not merge-ready. The 400-cycle TX preparation experiment survives scheduler bursts in simulation, but it incorrectly consumes a finite historical RX replay ring as future TX timing. That can produce NODATA, TxAlign jumps, clicks, and silence even while DMA/SYT remain alive.

The PR also identifies two recovery defects introduced by FW-64:

• AV/C overwrites DICE's single timing-loss callback.
• AV/C may restart transport without quiescing/re-priming CoreAudio's TX producer under the same epoch.

Validation

• ./build.sh --no-bump (Debug build)
• python3 tools/tx_data_horizon_burst_sim.py suite
• python3 tools/buffer_geometry_verify.py verify --max-io 1024
• git diff --check

The local Xcode scheme rewrite is intentionally excluded.