Add Mackie Onyx-i (Oxford) support: recognition, publication, asymmetric duplex streaming

[ottendorfcipher]

Adds support for the Mackie/LOUD Onyx-i mixers (Oxford OXFW971 production run). Developed and verified against an Onyx 820i behind an Apple TB3→TB2→FireWire adapter chain on macOS 26.5.2 / M3. Context and captures: #100.

Verified on hardware

• Recognition + CoreAudio publication: 8-in / 2-out, 24-bit, 44.1 kHz, blocking mode.
• Capture: signal verified through the device's main-mix FW pair; sustained sessions with zero geometry mismatches and zero rejected packets.
• Playback: sustained music playback; steady-state TX (constant producer lag, no deficit growth, no short-frame faults).
• Duplex sessions start and stop cleanly (FSM terminal Idle, status 0).
• The device stamps usable SYT, so the existing RX-SYT-authoritative clock model works unmodified.

Change summary (one commit per step)

1. feat(profiles) — recognition scaffold, sentinel-gated pending hardware capture.
2. feat(profiles) — recognition for the documented Onyx family ids from libffado 2.5.0 (Onyx-i Oxford 0x081216, Onyx 1640i 0x001640/0x000006, Blackbird 0x000007); identity only, same policy as the StudioLive siblings.
3. docs(profiles) — live Config-ROM capture: GUID 0x000FF20400003AFC, vendor 0x000FF2, model 0x081216, unit 0x00A02D/0x010001 (1394TA AV/C).
4. docs(profiles) — live unit-level 0xBF stream-format capture: capture 8ch MBLA, playback 2ch MBLA, compound AM824, 44.1/48/88.2/96 kHz, one isoch plug per direction, no MIDI.
5. feat(audio) — publication (M1): asymmetric ADK isoch profile (TX 2ch/DBS 2, RX 8ch/DBS 8, blocking, 44.1 kHz) dispatched ahead of the DICE GenericProfile() catch-all; LOUD vendor-wide forced blocking per snd-oxfw oxfw.c:189-196.
6. feat(avc) — publication bypass: the OXFW971's Music subunit implements no descriptor mechanism (standard descriptor access returns NOT IMPLEMENTED, the non-standard direct read also refuses; the parse yields 0 plugs), so the generic descriptor-driven path cannot publish this device. Publishes the profile-owned configuration from HandleInitializedUnit, same pattern as the BeBoB bypass.
7. feat(audio) — streaming (M2): MackieOnyxProtocol runtime duplex control; Onyx joins the IRM any-channel mask and the BeBoB start order (no pre-stream clock lock — the device is SYT-unaware per snd-oxfw); LOUD wrong-DBS RX stride quirk plumbed profile → coordinator → transport → consumer → RxAudioPacketProcessor (ALSA CIP_WRONG_DBS semantics, amdtp-stream.c:766-769); header dbs remains visible in telemetry.
8. fix(audio) — start-clock pin: a first-ever start has no session clock and the coordinator's default requests 48 kHz, which a 44.1-only device refuses at Prepare (kIOReturnUnsupported). Adds an Onyx entry to EffectiveStartClockForProfile, same mechanism as the Duet's pin.
9. fix(audio) — TxStreamPolicy.emptyPacketsDuringIdle = true: record-only sessions otherwise collapse with a monotonically growing TX prep deficit (observed 3.3k → 5.7k frames before session teardown, five consecutive cycles). Matches Phase88's policy on the same base.

Design decisions flagged for review

• MackieOnyxProtocol subclasses BeBoBProtocol. The base is protocol-neutral AV/C+CMP in practice: FCP dispatch, INPUT/OUTPUT PLUG SIGNAL FORMAT (0x18/0x19 — the same command snd-oxfw uses for OXFW rate control), CMP connect/verify/teardown, plug-0 streams, no-op mixer hook. If a second OXFW device adopts it, renaming the base to something family-neutral would be the follow-up.
• Rates are pinned to 44.1 kHz (the device's captured current rate) until an AV/C rate transition is validated against this hardware. The base's ApplyClockConfig already speaks the right command, so this should largely reduce to widening the rate list later; 88.2/96 kHz additionally sit behind AudioClockConfig's driver-wide gate.
• The model id 0x081216 is shared across the Oxford-run Onyx-i family; the static geometry here is 820i-verified. Siblings with different channel widths need a discovery-derived profile before they can stream (noted in the profile header; RegisterBeBoBProfile is the precedent).
• Tests: registry/factory integration-mode and identity pins; an asymmetry pin in DiceProfileTests locking 8-in/2-out against the Generic catch-all; an RX test pinning both the wrong-DBS quirk decode and the quirk-off failure mode (an oversized lying header decodes as zero events — indistinguishable from a silent device — rather than a geometry mismatch). The coordinator test mock is updated for the widened transport signatures.

References (Linux snd-oxfw/snd-dice/bebob, libffado 2.5.0, snd-firewire-ctl-services) were used as behavioral sources only, per the repository's reference policy; citations are in code comments. Local version-bump commits are excluded from this branch.

Not included

• Sample rates beyond 44.1 kHz (see above).
• Two resilience issues observed during bring-up that are independent of this device support — summarized in Mackie Onyx 820i (Loud 0x000ff2): live capture — Oxford-run unit reports shared Onyx-i id 0x081216, enumerates cleanly on Tahoe #100, reproducible with logs available: idle Thunderbolt adapter drop/re-attach cycles, and the audio endpoint not re-arming after such a cycle (subsequent StartIO returns kIOReturnNotReady until a fresh boot). Happy to file both as separate issues.

[ottendorfcipher]

Pushed three follow-up commits from continued hardware testing on the same 820i:

1. 6d70bd8 — safety-offset tuning: the Duet placeholder values left ~1.45 ms of TX headroom against observed host callback spikes of ~922 µs; widened per the measurements in the commit message.
2. 2ba3f97 — a 48 kHz rate-widening attempt (kept in history because its failure documented two engine findings).
3. dd899ac — its revert, after root-causing the regression it exposed: AudioDuplexCoordinator::RequestClockConfig persists session.pendingClock before executing the request, and the failure path (FailPendingClockRequest → clockRequests_.FailPending) completes the request bookkeeping without scrubbing the session. A rate change that fails at the ADK transport reconfig — which is currently unimplemented for AV/C static-profile devices (HandleChangeSampleRate → transport reconfig failed: kIOReturnUnsupported) — therefore leaves pendingClock=48000 armed; the next session start consumes it and programs the device to 48 kHz via SIGNAL FORMAT while the host graph runs 44.1, producing periodic audible dropouts ([TxExposure] reason=rate-mismatch, ~7,000 ppm sawtooth). Details and the defense-in-depth rationale are in the commit message.

Net state of the branch: 44.1 kHz-only, validated full duplex on hardware (capture and playback, including sustained mic use). The rate widening should return together with (a) the pending-clock hygiene fix and (b) AV/C rate-change support in the ADK reconfig path.

I can file the pending-clock hygiene defect and the reconfig gap as separate issues (plus the two bring-up resilience observations already summarized in #100) if that is useful.

[ottendorfcipher]

Adds support for the Mackie/LOUD Onyx-i mixers (Oxford OXFW971 production run). Developed and verified against an Onyx 820i behind an Apple TB3→TB2→FireWire adapter chain on macOS 26.5.2 / M3. Context and captures: #100.

Verified on hardware

• Recognition + CoreAudio publication: 8-in / 2-out, 24-bit, 44.1 kHz, blocking mode.
• Capture: signal verified through the device's main-mix FW pair; sustained sessions with zero geometry mismatches and zero rejected packets.
• Playback: sustained music playback; steady-state TX (constant producer lag, no deficit growth, no short-frame faults).
• Duplex sessions start and stop cleanly (FSM terminal Idle, status 0).
• The device stamps usable SYT, so the existing RX-SYT-authoritative clock model works unmodified.

Change summary (one commit per step)

1. feat(profiles) — recognition scaffold, sentinel-gated pending hardware capture.
2. feat(profiles) — recognition for the documented Onyx family ids from libffado 2.5.0 (Onyx-i Oxford 0x081216, Onyx 1640i 0x001640/0x000006, Blackbird 0x000007); identity only, same policy as the StudioLive siblings.
3. docs(profiles) — live Config-ROM capture: GUID 0x000FF20400003AFC, vendor 0x000FF2, model 0x081216, unit 0x00A02D/0x010001 (1394TA AV/C).
4. docs(profiles) — live unit-level 0xBF stream-format capture: capture 8ch MBLA, playback 2ch MBLA, compound AM824, 44.1/48/88.2/96 kHz, one isoch plug per direction, no MIDI.
5. feat(audio) — publication (M1): asymmetric ADK isoch profile (TX 2ch/DBS 2, RX 8ch/DBS 8, blocking, 44.1 kHz) dispatched ahead of the DICE GenericProfile() catch-all; LOUD vendor-wide forced blocking per snd-oxfw oxfw.c:189-196.
6. feat(avc) — publication bypass: the OXFW971's Music subunit implements no descriptor mechanism (standard descriptor access returns NOT IMPLEMENTED, the non-standard direct read also refuses; the parse yields 0 plugs), so the generic descriptor-driven path cannot publish this device. Publishes the profile-owned configuration from HandleInitializedUnit, same pattern as the BeBoB bypass.
7. feat(audio) — streaming (M2): MackieOnyxProtocol runtime duplex control; Onyx joins the IRM any-channel mask and the BeBoB start order (no pre-stream clock lock — the device is SYT-unaware per snd-oxfw); LOUD wrong-DBS RX stride quirk plumbed profile → coordinator → transport → consumer → RxAudioPacketProcessor (ALSA CIP_WRONG_DBS semantics, amdtp-stream.c:766-769); header dbs remains visible in telemetry.
8. fix(audio) — start-clock pin: a first-ever start has no session clock and the coordinator's default requests 48 kHz, which a 44.1-only device refuses at Prepare (kIOReturnUnsupported). Adds an Onyx entry to EffectiveStartClockForProfile, same mechanism as the Duet's pin.
9. fix(audio) — TxStreamPolicy.emptyPacketsDuringIdle = true: record-only sessions otherwise collapse with a monotonically growing TX prep deficit (observed 3.3k → 5.7k frames before session teardown, five consecutive cycles). Matches Phase88's policy on the same base.

Design decisions flagged for review

• MackieOnyxProtocol subclasses BeBoBProtocol. The base is protocol-neutral AV/C+CMP in practice: FCP dispatch, INPUT/OUTPUT PLUG SIGNAL FORMAT (0x18/0x19 — the same command snd-oxfw uses for OXFW rate control), CMP connect/verify/teardown, plug-0 streams, no-op mixer hook. If a second OXFW device adopts it, renaming the base to something family-neutral would be the follow-up.
• Rates are pinned to 44.1 kHz (the device's captured current rate) until an AV/C rate transition is validated against this hardware. The base's ApplyClockConfig already speaks the right command, so this should largely reduce to widening the rate list later; 88.2/96 kHz additionally sit behind AudioClockConfig's driver-wide gate.
• The model id 0x081216 is shared across the Oxford-run Onyx-i family; the static geometry here is 820i-verified. Siblings with different channel widths need a discovery-derived profile before they can stream (noted in the profile header; RegisterBeBoBProfile is the precedent).
• Tests: registry/factory integration-mode and identity pins; an asymmetry pin in DiceProfileTests locking 8-in/2-out against the Generic catch-all; an RX test pinning both the wrong-DBS quirk decode and the quirk-off failure mode (an oversized lying header decodes as zero events — indistinguishable from a silent device — rather than a geometry mismatch). The coordinator test mock is updated for the widened transport signatures.

References (Linux snd-oxfw/snd-dice/bebob, libffado 2.5.0, snd-firewire-ctl-services) were used as behavioral sources only, per the repository's reference policy; citations are in code comments. Local version-bump commits are excluded from this branch.

Not included

• Sample rates beyond 44.1 kHz (see above).
• Two resilience issues observed during bring-up that are independent of this device support — summarized in Mackie Onyx 820i (Loud 0x000ff2): live capture — Oxford-run unit reports shared Onyx-i id 0x081216, enumerates cleanly on Tahoe #100, reproducible with logs available: idle Thunderbolt adapter drop/re-attach cycles, and the audio endpoint not re-arming after such a cycle (subsequent StartIO returns kIOReturnNotReady until a fresh boot). Happy to file both as separate issues.

[ottendorfcipher]

Pushed three follow-up commits from continued hardware testing on the same 820i:

1. 6d70bd8 — safety-offset tuning: the Duet placeholder values left ~1.45 ms of TX headroom against observed host callback spikes of ~922 µs; widened per the measurements in the commit message.
2. 2ba3f97 — a 48 kHz rate-widening attempt (kept in history because its failure documented two engine findings).
3. dd899ac — its revert, after root-causing the regression it exposed: AudioDuplexCoordinator::RequestClockConfig persists session.pendingClock before executing the request, and the failure path (FailPendingClockRequest → clockRequests_.FailPending) completes the request bookkeeping without scrubbing the session. A rate change that fails at the ADK transport reconfig — which is currently unimplemented for AV/C static-profile devices (HandleChangeSampleRate → transport reconfig failed: kIOReturnUnsupported) — therefore leaves pendingClock=48000 armed; the next session start consumes it and programs the device to 48 kHz via SIGNAL FORMAT while the host graph runs 44.1, producing periodic audible dropouts ([TxExposure] reason=rate-mismatch, ~7,000 ppm sawtooth). Details and the defense-in-depth rationale are in the commit message.

Net state of the branch: 44.1 kHz-only, validated full duplex on hardware (capture and playback, including sustained mic use). The rate widening should return together with (a) the pending-clock hygiene fix and (b) AV/C rate-change support in the ADK reconfig path.

I can file the pending-clock hygiene defect and the reconfig gap as separate issues (plus the two bring-up resilience observations already summarized in #100) if that is useful.

[ottendorfcipher]

Adds support for the Mackie/LOUD Onyx-i mixers (Oxford OXFW971 production run). Developed and verified against an Onyx 820i behind an Apple TB3→TB2→FireWire adapter chain on macOS 26.5.2 / M3. Context and captures: #100.

Verified on hardware

• Recognition + CoreAudio publication: 8-in / 2-out, 24-bit, 44.1 kHz, blocking mode.
• Capture: signal verified through the device's main-mix FW pair; sustained sessions with zero geometry mismatches and zero rejected packets.
• Playback: sustained music playback; steady-state TX (constant producer lag, no deficit growth, no short-frame faults).
• Duplex sessions start and stop cleanly (FSM terminal Idle, status 0).
• The device stamps usable SYT, so the existing RX-SYT-authoritative clock model works unmodified.

Change summary (one commit per step)

1. feat(profiles) — recognition scaffold, sentinel-gated pending hardware capture.
2. feat(profiles) — recognition for the documented Onyx family ids from libffado 2.5.0 (Onyx-i Oxford 0x081216, Onyx 1640i 0x001640/0x000006, Blackbird 0x000007); identity only, same policy as the StudioLive siblings.
3. docs(profiles) — live Config-ROM capture: GUID 0x000FF20400003AFC, vendor 0x000FF2, model 0x081216, unit 0x00A02D/0x010001 (1394TA AV/C).
4. docs(profiles) — live unit-level 0xBF stream-format capture: capture 8ch MBLA, playback 2ch MBLA, compound AM824, 44.1/48/88.2/96 kHz, one isoch plug per direction, no MIDI.
5. feat(audio) — publication (M1): asymmetric ADK isoch profile (TX 2ch/DBS 2, RX 8ch/DBS 8, blocking, 44.1 kHz) dispatched ahead of the DICE GenericProfile() catch-all; LOUD vendor-wide forced blocking per snd-oxfw oxfw.c:189-196.
6. feat(avc) — publication bypass: the OXFW971's Music subunit implements no descriptor mechanism (standard descriptor access returns NOT IMPLEMENTED, the non-standard direct read also refuses; the parse yields 0 plugs), so the generic descriptor-driven path cannot publish this device. Publishes the profile-owned configuration from HandleInitializedUnit, same pattern as the BeBoB bypass.
7. feat(audio) — streaming (M2): MackieOnyxProtocol runtime duplex control; Onyx joins the IRM any-channel mask and the BeBoB start order (no pre-stream clock lock — the device is SYT-unaware per snd-oxfw); LOUD wrong-DBS RX stride quirk plumbed profile → coordinator → transport → consumer → RxAudioPacketProcessor (ALSA CIP_WRONG_DBS semantics, amdtp-stream.c:766-769); header dbs remains visible in telemetry.
8. fix(audio) — start-clock pin: a first-ever start has no session clock and the coordinator's default requests 48 kHz, which a 44.1-only device refuses at Prepare (kIOReturnUnsupported). Adds an Onyx entry to EffectiveStartClockForProfile, same mechanism as the Duet's pin.
9. fix(audio) — TxStreamPolicy.emptyPacketsDuringIdle = true: record-only sessions otherwise collapse with a monotonically growing TX prep deficit (observed 3.3k → 5.7k frames before session teardown, five consecutive cycles). Matches Phase88's policy on the same base.

Design decisions flagged for review

• MackieOnyxProtocol subclasses BeBoBProtocol. The base is protocol-neutral AV/C+CMP in practice: FCP dispatch, INPUT/OUTPUT PLUG SIGNAL FORMAT (0x18/0x19 — the same command snd-oxfw uses for OXFW rate control), CMP connect/verify/teardown, plug-0 streams, no-op mixer hook. If a second OXFW device adopts it, renaming the base to something family-neutral would be the follow-up.
• Rates are pinned to 44.1 kHz (the device's captured current rate) until an AV/C rate transition is validated against this hardware. The base's ApplyClockConfig already speaks the right command, so this should largely reduce to widening the rate list later; 88.2/96 kHz additionally sit behind AudioClockConfig's driver-wide gate.
• The model id 0x081216 is shared across the Oxford-run Onyx-i family; the static geometry here is 820i-verified. Siblings with different channel widths need a discovery-derived profile before they can stream (noted in the profile header; RegisterBeBoBProfile is the precedent).
• Tests: registry/factory integration-mode and identity pins; an asymmetry pin in DiceProfileTests locking 8-in/2-out against the Generic catch-all; an RX test pinning both the wrong-DBS quirk decode and the quirk-off failure mode (an oversized lying header decodes as zero events — indistinguishable from a silent device — rather than a geometry mismatch). The coordinator test mock is updated for the widened transport signatures.

References (Linux snd-oxfw/snd-dice/bebob, libffado 2.5.0, snd-firewire-ctl-services) were used as behavioral sources only, per the repository's reference policy; citations are in code comments. Local version-bump commits are excluded from this branch.

Not included

• Sample rates beyond 44.1 kHz (see above).
• Two resilience issues observed during bring-up that are independent of this device support — summarized in Mackie Onyx 820i (Loud 0x000ff2): live capture — Oxford-run unit reports shared Onyx-i id 0x081216, enumerates cleanly on Tahoe #100, reproducible with logs available: idle Thunderbolt adapter drop/re-attach cycles, and the audio endpoint not re-arming after such a cycle (subsequent StartIO returns kIOReturnNotReady until a fresh boot). Happy to file both as separate issues.

[ottendorfcipher]

Pushed three follow-up commits from continued hardware testing on the same 820i:

1. 6d70bd8 — safety-offset tuning: the Duet placeholder values left ~1.45 ms of TX headroom against observed host callback spikes of ~922 µs; widened per the measurements in the commit message.
2. 2ba3f97 — a 48 kHz rate-widening attempt (kept in history because its failure documented two engine findings).
3. dd899ac — its revert, after root-causing the regression it exposed: AudioDuplexCoordinator::RequestClockConfig persists session.pendingClock before executing the request, and the failure path (FailPendingClockRequest → clockRequests_.FailPending) completes the request bookkeeping without scrubbing the session. A rate change that fails at the ADK transport reconfig — which is currently unimplemented for AV/C static-profile devices (HandleChangeSampleRate → transport reconfig failed: kIOReturnUnsupported) — therefore leaves pendingClock=48000 armed; the next session start consumes it and programs the device to 48 kHz via SIGNAL FORMAT while the host graph runs 44.1, producing periodic audible dropouts ([TxExposure] reason=rate-mismatch, ~7,000 ppm sawtooth). Details and the defense-in-depth rationale are in the commit message.

Net state of the branch: 44.1 kHz-only, validated full duplex on hardware (capture and playback, including sustained mic use). The rate widening should return together with (a) the pending-clock hygiene fix and (b) AV/C rate-change support in the ADK reconfig path.

I can file the pending-clock hygiene defect and the reconfig gap as separate issues (plus the two bring-up resilience observations already summarized in #100) if that is useful.

[ottendorfcipher]

Adds support for the Mackie/LOUD Onyx-i mixers (Oxford OXFW971 production run). Developed and verified against an Onyx 820i behind an Apple TB3→TB2→FireWire adapter chain on macOS 26.5.2 / M3. Context and captures: #100.

Verified on hardware

• Recognition + CoreAudio publication: 8-in / 2-out, 24-bit, 44.1 kHz, blocking mode.
• Capture: signal verified through the device's main-mix FW pair; sustained sessions with zero geometry mismatches and zero rejected packets.
• Playback: sustained music playback; steady-state TX (constant producer lag, no deficit growth, no short-frame faults).
• Duplex sessions start and stop cleanly (FSM terminal Idle, status 0).
• The device stamps usable SYT, so the existing RX-SYT-authoritative clock model works unmodified.

Change summary (one commit per step)

1. feat(profiles) — recognition scaffold, sentinel-gated pending hardware capture.
2. feat(profiles) — recognition for the documented Onyx family ids from libffado 2.5.0 (Onyx-i Oxford 0x081216, Onyx 1640i 0x001640/0x000006, Blackbird 0x000007); identity only, same policy as the StudioLive siblings.
3. docs(profiles) — live Config-ROM capture: GUID 0x000FF20400003AFC, vendor 0x000FF2, model 0x081216, unit 0x00A02D/0x010001 (1394TA AV/C).
4. docs(profiles) — live unit-level 0xBF stream-format capture: capture 8ch MBLA, playback 2ch MBLA, compound AM824, 44.1/48/88.2/96 kHz, one isoch plug per direction, no MIDI.
5. feat(audio) — publication (M1): asymmetric ADK isoch profile (TX 2ch/DBS 2, RX 8ch/DBS 8, blocking, 44.1 kHz) dispatched ahead of the DICE GenericProfile() catch-all; LOUD vendor-wide forced blocking per snd-oxfw oxfw.c:189-196.
6. feat(avc) — publication bypass: the OXFW971's Music subunit implements no descriptor mechanism (standard descriptor access returns NOT IMPLEMENTED, the non-standard direct read also refuses; the parse yields 0 plugs), so the generic descriptor-driven path cannot publish this device. Publishes the profile-owned configuration from HandleInitializedUnit, same pattern as the BeBoB bypass.
7. feat(audio) — streaming (M2): MackieOnyxProtocol runtime duplex control; Onyx joins the IRM any-channel mask and the BeBoB start order (no pre-stream clock lock — the device is SYT-unaware per snd-oxfw); LOUD wrong-DBS RX stride quirk plumbed profile → coordinator → transport → consumer → RxAudioPacketProcessor (ALSA CIP_WRONG_DBS semantics, amdtp-stream.c:766-769); header dbs remains visible in telemetry.
8. fix(audio) — start-clock pin: a first-ever start has no session clock and the coordinator's default requests 48 kHz, which a 44.1-only device refuses at Prepare (kIOReturnUnsupported). Adds an Onyx entry to EffectiveStartClockForProfile, same mechanism as the Duet's pin.
9. fix(audio) — TxStreamPolicy.emptyPacketsDuringIdle = true: record-only sessions otherwise collapse with a monotonically growing TX prep deficit (observed 3.3k → 5.7k frames before session teardown, five consecutive cycles). Matches Phase88's policy on the same base.

Design decisions flagged for review

• MackieOnyxProtocol subclasses BeBoBProtocol. The base is protocol-neutral AV/C+CMP in practice: FCP dispatch, INPUT/OUTPUT PLUG SIGNAL FORMAT (0x18/0x19 — the same command snd-oxfw uses for OXFW rate control), CMP connect/verify/teardown, plug-0 streams, no-op mixer hook. If a second OXFW device adopts it, renaming the base to something family-neutral would be the follow-up.
• Rates are pinned to 44.1 kHz (the device's captured current rate) until an AV/C rate transition is validated against this hardware. The base's ApplyClockConfig already speaks the right command, so this should largely reduce to widening the rate list later; 88.2/96 kHz additionally sit behind AudioClockConfig's driver-wide gate.
• The model id 0x081216 is shared across the Oxford-run Onyx-i family; the static geometry here is 820i-verified. Siblings with different channel widths need a discovery-derived profile before they can stream (noted in the profile header; RegisterBeBoBProfile is the precedent).
• Tests: registry/factory integration-mode and identity pins; an asymmetry pin in DiceProfileTests locking 8-in/2-out against the Generic catch-all; an RX test pinning both the wrong-DBS quirk decode and the quirk-off failure mode (an oversized lying header decodes as zero events — indistinguishable from a silent device — rather than a geometry mismatch). The coordinator test mock is updated for the widened transport signatures.

References (Linux snd-oxfw/snd-dice/bebob, libffado 2.5.0, snd-firewire-ctl-services) were used as behavioral sources only, per the repository's reference policy; citations are in code comments. Local version-bump commits are excluded from this branch.

Not included

• Sample rates beyond 44.1 kHz (see above).
• Two resilience issues observed during bring-up that are independent of this device support — summarized in Mackie Onyx 820i (Loud 0x000ff2): live capture — Oxford-run unit reports shared Onyx-i id 0x081216, enumerates cleanly on Tahoe #100, reproducible with logs available: idle Thunderbolt adapter drop/re-attach cycles, and the audio endpoint not re-arming after such a cycle (subsequent StartIO returns kIOReturnNotReady until a fresh boot). Happy to file both as separate issues.

[ottendorfcipher]

Pushed three follow-up commits from continued hardware testing on the same 820i:

1. 6d70bd8 — safety-offset tuning: the Duet placeholder values left ~1.45 ms of TX headroom against observed host callback spikes of ~922 µs; widened per the measurements in the commit message.
2. 2ba3f97 — a 48 kHz rate-widening attempt (kept in history because its failure documented two engine findings).
3. dd899ac — its revert, after root-causing the regression it exposed: AudioDuplexCoordinator::RequestClockConfig persists session.pendingClock before executing the request, and the failure path (FailPendingClockRequest → clockRequests_.FailPending) completes the request bookkeeping without scrubbing the session. A rate change that fails at the ADK transport reconfig — which is currently unimplemented for AV/C static-profile devices (HandleChangeSampleRate → transport reconfig failed: kIOReturnUnsupported) — therefore leaves pendingClock=48000 armed; the next session start consumes it and programs the device to 48 kHz via SIGNAL FORMAT while the host graph runs 44.1, producing periodic audible dropouts ([TxExposure] reason=rate-mismatch, ~7,000 ppm sawtooth). Details and the defense-in-depth rationale are in the commit message.

Net state of the branch: 44.1 kHz-only, validated full duplex on hardware (capture and playback, including sustained mic use). The rate widening should return together with (a) the pending-clock hygiene fix and (b) AV/C rate-change support in the ADK reconfig path.

I can file the pending-clock hygiene defect and the reconfig gap as separate issues (plus the two bring-up resilience observations already summarized in #100) if that is useful.

[ottendorfcipher]

Adds support for the Mackie/LOUD Onyx-i mixers (Oxford OXFW971 production run). Developed and verified against an Onyx 820i behind an Apple TB3→TB2→FireWire adapter chain on macOS 26.5.2 / M3. Context and captures: #100.

Verified on hardware

• Recognition + CoreAudio publication: 8-in / 2-out, 24-bit, 44.1 kHz, blocking mode.
• Capture: signal verified through the device's main-mix FW pair; sustained sessions with zero geometry mismatches and zero rejected packets.
• Playback: sustained music playback; steady-state TX (constant producer lag, no deficit growth, no short-frame faults).
• Duplex sessions start and stop cleanly (FSM terminal Idle, status 0).
• The device stamps usable SYT, so the existing RX-SYT-authoritative clock model works unmodified.

Change summary (one commit per step)

1. feat(profiles) — recognition scaffold, sentinel-gated pending hardware capture.
2. feat(profiles) — recognition for the documented Onyx family ids from libffado 2.5.0 (Onyx-i Oxford 0x081216, Onyx 1640i 0x001640/0x000006, Blackbird 0x000007); identity only, same policy as the StudioLive siblings.
3. docs(profiles) — live Config-ROM capture: GUID 0x000FF20400003AFC, vendor 0x000FF2, model 0x081216, unit 0x00A02D/0x010001 (1394TA AV/C).
4. docs(profiles) — live unit-level 0xBF stream-format capture: capture 8ch MBLA, playback 2ch MBLA, compound AM824, 44.1/48/88.2/96 kHz, one isoch plug per direction, no MIDI.
5. feat(audio) — publication (M1): asymmetric ADK isoch profile (TX 2ch/DBS 2, RX 8ch/DBS 8, blocking, 44.1 kHz) dispatched ahead of the DICE GenericProfile() catch-all; LOUD vendor-wide forced blocking per snd-oxfw oxfw.c:189-196.
6. feat(avc) — publication bypass: the OXFW971's Music subunit implements no descriptor mechanism (standard descriptor access returns NOT IMPLEMENTED, the non-standard direct read also refuses; the parse yields 0 plugs), so the generic descriptor-driven path cannot publish this device. Publishes the profile-owned configuration from HandleInitializedUnit, same pattern as the BeBoB bypass.
7. feat(audio) — streaming (M2): MackieOnyxProtocol runtime duplex control; Onyx joins the IRM any-channel mask and the BeBoB start order (no pre-stream clock lock — the device is SYT-unaware per snd-oxfw); LOUD wrong-DBS RX stride quirk plumbed profile → coordinator → transport → consumer → RxAudioPacketProcessor (ALSA CIP_WRONG_DBS semantics, amdtp-stream.c:766-769); header dbs remains visible in telemetry.
8. fix(audio) — start-clock pin: a first-ever start has no session clock and the coordinator's default requests 48 kHz, which a 44.1-only device refuses at Prepare (kIOReturnUnsupported). Adds an Onyx entry to EffectiveStartClockForProfile, same mechanism as the Duet's pin.
9. fix(audio) — TxStreamPolicy.emptyPacketsDuringIdle = true: record-only sessions otherwise collapse with a monotonically growing TX prep deficit (observed 3.3k → 5.7k frames before session teardown, five consecutive cycles). Matches Phase88's policy on the same base.

Design decisions flagged for review

• MackieOnyxProtocol subclasses BeBoBProtocol. The base is protocol-neutral AV/C+CMP in practice: FCP dispatch, INPUT/OUTPUT PLUG SIGNAL FORMAT (0x18/0x19 — the same command snd-oxfw uses for OXFW rate control), CMP connect/verify/teardown, plug-0 streams, no-op mixer hook. If a second OXFW device adopts it, renaming the base to something family-neutral would be the follow-up.
• Rates are pinned to 44.1 kHz (the device's captured current rate) until an AV/C rate transition is validated against this hardware. The base's ApplyClockConfig already speaks the right command, so this should largely reduce to widening the rate list later; 88.2/96 kHz additionally sit behind AudioClockConfig's driver-wide gate.
• The model id 0x081216 is shared across the Oxford-run Onyx-i family; the static geometry here is 820i-verified. Siblings with different channel widths need a discovery-derived profile before they can stream (noted in the profile header; RegisterBeBoBProfile is the precedent).
• Tests: registry/factory integration-mode and identity pins; an asymmetry pin in DiceProfileTests locking 8-in/2-out against the Generic catch-all; an RX test pinning both the wrong-DBS quirk decode and the quirk-off failure mode (an oversized lying header decodes as zero events — indistinguishable from a silent device — rather than a geometry mismatch). The coordinator test mock is updated for the widened transport signatures.

References (Linux snd-oxfw/snd-dice/bebob, libffado 2.5.0, snd-firewire-ctl-services) were used as behavioral sources only, per the repository's reference policy; citations are in code comments. Local version-bump commits are excluded from this branch.

Not included

• Sample rates beyond 44.1 kHz (see above).
• Two resilience issues observed during bring-up that are independent of this device support — summarized in Mackie Onyx 820i (Loud 0x000ff2): live capture — Oxford-run unit reports shared Onyx-i id 0x081216, enumerates cleanly on Tahoe #100, reproducible with logs available: idle Thunderbolt adapter drop/re-attach cycles, and the audio endpoint not re-arming after such a cycle (subsequent StartIO returns kIOReturnNotReady until a fresh boot). Happy to file both as separate issues.

[ottendorfcipher]

Pushed three follow-up commits from continued hardware testing on the same 820i:

1. 6d70bd8 — safety-offset tuning: the Duet placeholder values left ~1.45 ms of TX headroom against observed host callback spikes of ~922 µs; widened per the measurements in the commit message.
2. 2ba3f97 — a 48 kHz rate-widening attempt (kept in history because its failure documented two engine findings).
3. dd899ac — its revert, after root-causing the regression it exposed: AudioDuplexCoordinator::RequestClockConfig persists session.pendingClock before executing the request, and the failure path (FailPendingClockRequest → clockRequests_.FailPending) completes the request bookkeeping without scrubbing the session. A rate change that fails at the ADK transport reconfig — which is currently unimplemented for AV/C static-profile devices (HandleChangeSampleRate → transport reconfig failed: kIOReturnUnsupported) — therefore leaves pendingClock=48000 armed; the next session start consumes it and programs the device to 48 kHz via SIGNAL FORMAT while the host graph runs 44.1, producing periodic audible dropouts ([TxExposure] reason=rate-mismatch, ~7,000 ppm sawtooth). Details and the defense-in-depth rationale are in the commit message.

Net state of the branch: 44.1 kHz-only, validated full duplex on hardware (capture and playback, including sustained mic use). The rate widening should return together with (a) the pending-clock hygiene fix and (b) AV/C rate-change support in the ADK reconfig path.

I can file the pending-clock hygiene defect and the reconfig gap as separate issues (plus the two bring-up resilience observations already summarized in #100) if that is useful.

[ottendorfcipher]

Adds support for the Mackie/LOUD Onyx-i mixers (Oxford OXFW971 production run). Developed and verified against an Onyx 820i behind an Apple TB3→TB2→FireWire adapter chain on macOS 26.5.2 / M3. Context and captures: #100.

Verified on hardware

• Recognition + CoreAudio publication: 8-in / 2-out, 24-bit, 44.1 kHz, blocking mode.
• Capture: signal verified through the device's main-mix FW pair; sustained sessions with zero geometry mismatches and zero rejected packets.
• Playback: sustained music playback; steady-state TX (constant producer lag, no deficit growth, no short-frame faults).
• Duplex sessions start and stop cleanly (FSM terminal Idle, status 0).
• The device stamps usable SYT, so the existing RX-SYT-authoritative clock model works unmodified.

Change summary (one commit per step)

1. feat(profiles) — recognition scaffold, sentinel-gated pending hardware capture.
2. feat(profiles) — recognition for the documented Onyx family ids from libffado 2.5.0 (Onyx-i Oxford 0x081216, Onyx 1640i 0x001640/0x000006, Blackbird 0x000007); identity only, same policy as the StudioLive siblings.
3. docs(profiles) — live Config-ROM capture: GUID 0x000FF20400003AFC, vendor 0x000FF2, model 0x081216, unit 0x00A02D/0x010001 (1394TA AV/C).
4. docs(profiles) — live unit-level 0xBF stream-format capture: capture 8ch MBLA, playback 2ch MBLA, compound AM824, 44.1/48/88.2/96 kHz, one isoch plug per direction, no MIDI.
5. feat(audio) — publication (M1): asymmetric ADK isoch profile (TX 2ch/DBS 2, RX 8ch/DBS 8, blocking, 44.1 kHz) dispatched ahead of the DICE GenericProfile() catch-all; LOUD vendor-wide forced blocking per snd-oxfw oxfw.c:189-196.
6. feat(avc) — publication bypass: the OXFW971's Music subunit implements no descriptor mechanism (standard descriptor access returns NOT IMPLEMENTED, the non-standard direct read also refuses; the parse yields 0 plugs), so the generic descriptor-driven path cannot publish this device. Publishes the profile-owned configuration from HandleInitializedUnit, same pattern as the BeBoB bypass.
7. feat(audio) — streaming (M2): MackieOnyxProtocol runtime duplex control; Onyx joins the IRM any-channel mask and the BeBoB start order (no pre-stream clock lock — the device is SYT-unaware per snd-oxfw); LOUD wrong-DBS RX stride quirk plumbed profile → coordinator → transport → consumer → RxAudioPacketProcessor (ALSA CIP_WRONG_DBS semantics, amdtp-stream.c:766-769); header dbs remains visible in telemetry.
8. fix(audio) — start-clock pin: a first-ever start has no session clock and the coordinator's default requests 48 kHz, which a 44.1-only device refuses at Prepare (kIOReturnUnsupported). Adds an Onyx entry to EffectiveStartClockForProfile, same mechanism as the Duet's pin.
9. fix(audio) — TxStreamPolicy.emptyPacketsDuringIdle = true: record-only sessions otherwise collapse with a monotonically growing TX prep deficit (observed 3.3k → 5.7k frames before session teardown, five consecutive cycles). Matches Phase88's policy on the same base.

Design decisions flagged for review

• MackieOnyxProtocol subclasses BeBoBProtocol. The base is protocol-neutral AV/C+CMP in practice: FCP dispatch, INPUT/OUTPUT PLUG SIGNAL FORMAT (0x18/0x19 — the same command snd-oxfw uses for OXFW rate control), CMP connect/verify/teardown, plug-0 streams, no-op mixer hook. If a second OXFW device adopts it, renaming the base to something family-neutral would be the follow-up.
• Rates are pinned to 44.1 kHz (the device's captured current rate) until an AV/C rate transition is validated against this hardware. The base's ApplyClockConfig already speaks the right command, so this should largely reduce to widening the rate list later; 88.2/96 kHz additionally sit behind AudioClockConfig's driver-wide gate.
• The model id 0x081216 is shared across the Oxford-run Onyx-i family; the static geometry here is 820i-verified. Siblings with different channel widths need a discovery-derived profile before they can stream (noted in the profile header; RegisterBeBoBProfile is the precedent).
• Tests: registry/factory integration-mode and identity pins; an asymmetry pin in DiceProfileTests locking 8-in/2-out against the Generic catch-all; an RX test pinning both the wrong-DBS quirk decode and the quirk-off failure mode (an oversized lying header decodes as zero events — indistinguishable from a silent device — rather than a geometry mismatch). The coordinator test mock is updated for the widened transport signatures.

References (Linux snd-oxfw/snd-dice/bebob, libffado 2.5.0, snd-firewire-ctl-services) were used as behavioral sources only, per the repository's reference policy; citations are in code comments. Local version-bump commits are excluded from this branch.

Not included

• Sample rates beyond 44.1 kHz (see above).
• Two resilience issues observed during bring-up that are independent of this device support — summarized in Mackie Onyx 820i (Loud 0x000ff2): live capture — Oxford-run unit reports shared Onyx-i id 0x081216, enumerates cleanly on Tahoe #100, reproducible with logs available: idle Thunderbolt adapter drop/re-attach cycles, and the audio endpoint not re-arming after such a cycle (subsequent StartIO returns kIOReturnNotReady until a fresh boot). Happy to file both as separate issues.

[ottendorfcipher]

Pushed three follow-up commits from continued hardware testing on the same 820i:

1. 6d70bd8 — safety-offset tuning: the Duet placeholder values left ~1.45 ms of TX headroom against observed host callback spikes of ~922 µs; widened per the measurements in the commit message.
2. 2ba3f97 — a 48 kHz rate-widening attempt (kept in history because its failure documented two engine findings).
3. dd899ac — its revert, after root-causing the regression it exposed: AudioDuplexCoordinator::RequestClockConfig persists session.pendingClock before executing the request, and the failure path (FailPendingClockRequest → clockRequests_.FailPending) completes the request bookkeeping without scrubbing the session. A rate change that fails at the ADK transport reconfig — which is currently unimplemented for AV/C static-profile devices (HandleChangeSampleRate → transport reconfig failed: kIOReturnUnsupported) — therefore leaves pendingClock=48000 armed; the next session start consumes it and programs the device to 48 kHz via SIGNAL FORMAT while the host graph runs 44.1, producing periodic audible dropouts ([TxExposure] reason=rate-mismatch, ~7,000 ppm sawtooth). Details and the defense-in-depth rationale are in the commit message.

Net state of the branch: 44.1 kHz-only, validated full duplex on hardware (capture and playback, including sustained mic use). The rate widening should return together with (a) the pending-clock hygiene fix and (b) AV/C rate-change support in the ADK reconfig path.

I can file the pending-clock hygiene defect and the reconfig gap as separate issues (plus the two bring-up resilience observations already summarized in #100) if that is useful.

[ottendorfcipher]

Adds support for the Mackie/LOUD Onyx-i mixers (Oxford OXFW971 production run). Developed and verified against an Onyx 820i behind an Apple TB3→TB2→FireWire adapter chain on macOS 26.5.2 / M3. Context and captures: #100.

Verified on hardware

• Recognition + CoreAudio publication: 8-in / 2-out, 24-bit, 44.1 kHz, blocking mode.
• Capture: signal verified through the device's main-mix FW pair; sustained sessions with zero geometry mismatches and zero rejected packets.
• Playback: sustained music playback; steady-state TX (constant producer lag, no deficit growth, no short-frame faults).
• Duplex sessions start and stop cleanly (FSM terminal Idle, status 0).
• The device stamps usable SYT, so the existing RX-SYT-authoritative clock model works unmodified.

Change summary (one commit per step)

1. feat(profiles) — recognition scaffold, sentinel-gated pending hardware capture.
2. feat(profiles) — recognition for the documented Onyx family ids from libffado 2.5.0 (Onyx-i Oxford 0x081216, Onyx 1640i 0x001640/0x000006, Blackbird 0x000007); identity only, same policy as the StudioLive siblings.
3. docs(profiles) — live Config-ROM capture: GUID 0x000FF20400003AFC, vendor 0x000FF2, model 0x081216, unit 0x00A02D/0x010001 (1394TA AV/C).
4. docs(profiles) — live unit-level 0xBF stream-format capture: capture 8ch MBLA, playback 2ch MBLA, compound AM824, 44.1/48/88.2/96 kHz, one isoch plug per direction, no MIDI.
5. feat(audio) — publication (M1): asymmetric ADK isoch profile (TX 2ch/DBS 2, RX 8ch/DBS 8, blocking, 44.1 kHz) dispatched ahead of the DICE GenericProfile() catch-all; LOUD vendor-wide forced blocking per snd-oxfw oxfw.c:189-196.
6. feat(avc) — publication bypass: the OXFW971's Music subunit implements no descriptor mechanism (standard descriptor access returns NOT IMPLEMENTED, the non-standard direct read also refuses; the parse yields 0 plugs), so the generic descriptor-driven path cannot publish this device. Publishes the profile-owned configuration from HandleInitializedUnit, same pattern as the BeBoB bypass.
7. feat(audio) — streaming (M2): MackieOnyxProtocol runtime duplex control; Onyx joins the IRM any-channel mask and the BeBoB start order (no pre-stream clock lock — the device is SYT-unaware per snd-oxfw); LOUD wrong-DBS RX stride quirk plumbed profile → coordinator → transport → consumer → RxAudioPacketProcessor (ALSA CIP_WRONG_DBS semantics, amdtp-stream.c:766-769); header dbs remains visible in telemetry.
8. fix(audio) — start-clock pin: a first-ever start has no session clock and the coordinator's default requests 48 kHz, which a 44.1-only device refuses at Prepare (kIOReturnUnsupported). Adds an Onyx entry to EffectiveStartClockForProfile, same mechanism as the Duet's pin.
9. fix(audio) — TxStreamPolicy.emptyPacketsDuringIdle = true: record-only sessions otherwise collapse with a monotonically growing TX prep deficit (observed 3.3k → 5.7k frames before session teardown, five consecutive cycles). Matches Phase88's policy on the same base.

Design decisions flagged for review

• MackieOnyxProtocol subclasses BeBoBProtocol. The base is protocol-neutral AV/C+CMP in practice: FCP dispatch, INPUT/OUTPUT PLUG SIGNAL FORMAT (0x18/0x19 — the same command snd-oxfw uses for OXFW rate control), CMP connect/verify/teardown, plug-0 streams, no-op mixer hook. If a second OXFW device adopts it, renaming the base to something family-neutral would be the follow-up.
• Rates are pinned to 44.1 kHz (the device's captured current rate) until an AV/C rate transition is validated against this hardware. The base's ApplyClockConfig already speaks the right command, so this should largely reduce to widening the rate list later; 88.2/96 kHz additionally sit behind AudioClockConfig's driver-wide gate.
• The model id 0x081216 is shared across the Oxford-run Onyx-i family; the static geometry here is 820i-verified. Siblings with different channel widths need a discovery-derived profile before they can stream (noted in the profile header; RegisterBeBoBProfile is the precedent).
• Tests: registry/factory integration-mode and identity pins; an asymmetry pin in DiceProfileTests locking 8-in/2-out against the Generic catch-all; an RX test pinning both the wrong-DBS quirk decode and the quirk-off failure mode (an oversized lying header decodes as zero events — indistinguishable from a silent device — rather than a geometry mismatch). The coordinator test mock is updated for the widened transport signatures.

References (Linux snd-oxfw/snd-dice/bebob, libffado 2.5.0, snd-firewire-ctl-services) were used as behavioral sources only, per the repository's reference policy; citations are in code comments. Local version-bump commits are excluded from this branch.

Not included

• Sample rates beyond 44.1 kHz (see above).
• Two resilience issues observed during bring-up that are independent of this device support — summarized in Mackie Onyx 820i (Loud 0x000ff2): live capture — Oxford-run unit reports shared Onyx-i id 0x081216, enumerates cleanly on Tahoe #100, reproducible with logs available: idle Thunderbolt adapter drop/re-attach cycles, and the audio endpoint not re-arming after such a cycle (subsequent StartIO returns kIOReturnNotReady until a fresh boot). Happy to file both as separate issues.

[ottendorfcipher]

Pushed three follow-up commits from continued hardware testing on the same 820i:

1. 6d70bd8 — safety-offset tuning: the Duet placeholder values left ~1.45 ms of TX headroom against observed host callback spikes of ~922 µs; widened per the measurements in the commit message.
2. 2ba3f97 — a 48 kHz rate-widening attempt (kept in history because its failure documented two engine findings).
3. dd899ac — its revert, after root-causing the regression it exposed: AudioDuplexCoordinator::RequestClockConfig persists session.pendingClock before executing the request, and the failure path (FailPendingClockRequest → clockRequests_.FailPending) completes the request bookkeeping without scrubbing the session. A rate change that fails at the ADK transport reconfig — which is currently unimplemented for AV/C static-profile devices (HandleChangeSampleRate → transport reconfig failed: kIOReturnUnsupported) — therefore leaves pendingClock=48000 armed; the next session start consumes it and programs the device to 48 kHz via SIGNAL FORMAT while the host graph runs 44.1, producing periodic audible dropouts ([TxExposure] reason=rate-mismatch, ~7,000 ppm sawtooth). Details and the defense-in-depth rationale are in the commit message.

Net state of the branch: 44.1 kHz-only, validated full duplex on hardware (capture and playback, including sustained mic use). The rate widening should return together with (a) the pending-clock hygiene fix and (b) AV/C rate-change support in the ADK reconfig path.

I can file the pending-clock hygiene defect and the reconfig gap as separate issues (plus the two bring-up resilience observations already summarized in #100) if that is useful.

[ottendorfcipher]

Adds support for the Mackie/LOUD Onyx-i mixers (Oxford OXFW971 production run). Developed and verified against an Onyx 820i behind an Apple TB3→TB2→FireWire adapter chain on macOS 26.5.2 / M3. Context and captures: #100.

Verified on hardware

• Recognition + CoreAudio publication: 8-in / 2-out, 24-bit, 44.1 kHz, blocking mode.
• Capture: signal verified through the device's main-mix FW pair; sustained sessions with zero geometry mismatches and zero rejected packets.
• Playback: sustained music playback; steady-state TX (constant producer lag, no deficit growth, no short-frame faults).
• Duplex sessions start and stop cleanly (FSM terminal Idle, status 0).
• The device stamps usable SYT, so the existing RX-SYT-authoritative clock model works unmodified.

Change summary (one commit per step)

1. feat(profiles) — recognition scaffold, sentinel-gated pending hardware capture.
2. feat(profiles) — recognition for the documented Onyx family ids from libffado 2.5.0 (Onyx-i Oxford 0x081216, Onyx 1640i 0x001640/0x000006, Blackbird 0x000007); identity only, same policy as the StudioLive siblings.
3. docs(profiles) — live Config-ROM capture: GUID 0x000FF20400003AFC, vendor 0x000FF2, model 0x081216, unit 0x00A02D/0x010001 (1394TA AV/C).
4. docs(profiles) — live unit-level 0xBF stream-format capture: capture 8ch MBLA, playback 2ch MBLA, compound AM824, 44.1/48/88.2/96 kHz, one isoch plug per direction, no MIDI.
5. feat(audio) — publication (M1): asymmetric ADK isoch profile (TX 2ch/DBS 2, RX 8ch/DBS 8, blocking, 44.1 kHz) dispatched ahead of the DICE GenericProfile() catch-all; LOUD vendor-wide forced blocking per snd-oxfw oxfw.c:189-196.
6. feat(avc) — publication bypass: the OXFW971's Music subunit implements no descriptor mechanism (standard descriptor access returns NOT IMPLEMENTED, the non-standard direct read also refuses; the parse yields 0 plugs), so the generic descriptor-driven path cannot publish this device. Publishes the profile-owned configuration from HandleInitializedUnit, same pattern as the BeBoB bypass.
7. feat(audio) — streaming (M2): MackieOnyxProtocol runtime duplex control; Onyx joins the IRM any-channel mask and the BeBoB start order (no pre-stream clock lock — the device is SYT-unaware per snd-oxfw); LOUD wrong-DBS RX stride quirk plumbed profile → coordinator → transport → consumer → RxAudioPacketProcessor (ALSA CIP_WRONG_DBS semantics, amdtp-stream.c:766-769); header dbs remains visible in telemetry.
8. fix(audio) — start-clock pin: a first-ever start has no session clock and the coordinator's default requests 48 kHz, which a 44.1-only device refuses at Prepare (kIOReturnUnsupported). Adds an Onyx entry to EffectiveStartClockForProfile, same mechanism as the Duet's pin.
9. fix(audio) — TxStreamPolicy.emptyPacketsDuringIdle = true: record-only sessions otherwise collapse with a monotonically growing TX prep deficit (observed 3.3k → 5.7k frames before session teardown, five consecutive cycles). Matches Phase88's policy on the same base.

Design decisions flagged for review

• MackieOnyxProtocol subclasses BeBoBProtocol. The base is protocol-neutral AV/C+CMP in practice: FCP dispatch, INPUT/OUTPUT PLUG SIGNAL FORMAT (0x18/0x19 — the same command snd-oxfw uses for OXFW rate control), CMP connect/verify/teardown, plug-0 streams, no-op mixer hook. If a second OXFW device adopts it, renaming the base to something family-neutral would be the follow-up.
• Rates are pinned to 44.1 kHz (the device's captured current rate) until an AV/C rate transition is validated against this hardware. The base's ApplyClockConfig already speaks the right command, so this should largely reduce to widening the rate list later; 88.2/96 kHz additionally sit behind AudioClockConfig's driver-wide gate.
• The model id 0x081216 is shared across the Oxford-run Onyx-i family; the static geometry here is 820i-verified. Siblings with different channel widths need a discovery-derived profile before they can stream (noted in the profile header; RegisterBeBoBProfile is the precedent).
• Tests: registry/factory integration-mode and identity pins; an asymmetry pin in DiceProfileTests locking 8-in/2-out against the Generic catch-all; an RX test pinning both the wrong-DBS quirk decode and the quirk-off failure mode (an oversized lying header decodes as zero events — indistinguishable from a silent device — rather than a geometry mismatch). The coordinator test mock is updated for the widened transport signatures.

References (Linux snd-oxfw/snd-dice/bebob, libffado 2.5.0, snd-firewire-ctl-services) were used as behavioral sources only, per the repository's reference policy; citations are in code comments. Local version-bump commits are excluded from this branch.

Not included

• Sample rates beyond 44.1 kHz (see above).
• Two resilience issues observed during bring-up that are independent of this device support — summarized in Mackie Onyx 820i (Loud 0x000ff2): live capture — Oxford-run unit reports shared Onyx-i id 0x081216, enumerates cleanly on Tahoe #100, reproducible with logs available: idle Thunderbolt adapter drop/re-attach cycles, and the audio endpoint not re-arming after such a cycle (subsequent StartIO returns kIOReturnNotReady until a fresh boot). Happy to file both as separate issues.

[ottendorfcipher]

Pushed three follow-up commits from continued hardware testing on the same 820i:

1. 6d70bd8 — safety-offset tuning: the Duet placeholder values left ~1.45 ms of TX headroom against observed host callback spikes of ~922 µs; widened per the measurements in the commit message.
2. 2ba3f97 — a 48 kHz rate-widening attempt (kept in history because its failure documented two engine findings).
3. dd899ac — its revert, after root-causing the regression it exposed: AudioDuplexCoordinator::RequestClockConfig persists session.pendingClock before executing the request, and the failure path (FailPendingClockRequest → clockRequests_.FailPending) completes the request bookkeeping without scrubbing the session. A rate change that fails at the ADK transport reconfig — which is currently unimplemented for AV/C static-profile devices (HandleChangeSampleRate → transport reconfig failed: kIOReturnUnsupported) — therefore leaves pendingClock=48000 armed; the next session start consumes it and programs the device to 48 kHz via SIGNAL FORMAT while the host graph runs 44.1, producing periodic audible dropouts ([TxExposure] reason=rate-mismatch, ~7,000 ppm sawtooth). Details and the defense-in-depth rationale are in the commit message.

Net state of the branch: 44.1 kHz-only, validated full duplex on hardware (capture and playback, including sustained mic use). The rate widening should return together with (a) the pending-clock hygiene fix and (b) AV/C rate-change support in the ADK reconfig path.

I can file the pending-clock hygiene defect and the reconfig gap as separate issues (plus the two bring-up resilience observations already summarized in #100) if that is useful.