The engine's call to Music3SelectDepthArm is reachable but not gated, and #1131 no longer covers it. #1131 named two twins — the DiT arm's production switch and the depth decoder's — and PR #1821 closes the DiT half by running the shipped engine on thor:gpu0 inside an rc lease. The depth half is untouched by that change and needs its own issue rather than riding a closed one, per AGENTS.md ## Every change starts from an issue.
What is unreached. src/vllm/model_executor/models/minimax_music3_speech.cpp:638 calls Music3SelectDepthArm. Deleting that call leaves test_minimax_music3_ar 37/37 · 640/640 and test_minimax_music3_speech 9/9 · 223/223 green (.agents/specs/minimax-music3.md §19.5 carries the mutation and the binary hashes). The rule it calls is gated on both sides of its condition; what is owed is the call.
Why no existing gate sees it. Two reasons, and the second is the one that matters.
--speech-device 1 is refused by name on a CPU-only build before a queue exists, so no CI runner can take the branch.- The instrument that would answer it is not read from production.
ar.depth_staging is emitted at src/vllm/model_executor/models/minimax_music3_llm.cpp:582 and no test reads it. §19.6's "device path TAKEN" leg rides test_minimax_music3_ar, whose observable is Music3DepthDeviceForwardCount() — which §19.5 itself records as unreachable from production, its only readers being the tests written for it (tests/vllm/models/test_minimax_music3_ar.cpp:1325,1351,1583,1589,1753,1758). A counter no production run reads measures a class, not a capability.
It is closable by exactly #1821's method, which is why this is a scoped issue and not a research one. A tests/parity/test_minimax_music3_depth_arm_real.cpp entering through include/vllm.h (vllm_speech_engine_load with device = 1, then vllm_synthesize), labelled gpu;checkpoint;music3, exiting 77 without a device or checkpoint, asserting ar.depth_staging calls == 1 from the engine's own profile::Report with the host bucket absent. #1821's run already fired ar.depth_staging once on thor:gpu0 without asserting it, so the instrument is known to be live on the real path. The acceptance criterion is the same one line: with the engine's call to Music3SelectDepthArm deleted, the gate reds.
Owned by row MUSIC3-DEPTH-DEVICE (lane #672). Recorded under ## Owed in .agents/specs/minimax-music3.md §19.7, which this issue replaces #1131 as the tracker for.
The engine's call to
Music3SelectDepthArmis reachable but not gated, and #1131 no longer covers it. #1131 named two twins — the DiT arm's production switch and the depth decoder's — and PR #1821 closes the DiT half by running the shipped engine onthor:gpu0inside anrclease. The depth half is untouched by that change and needs its own issue rather than riding a closed one, perAGENTS.md## Every change starts from an issue.What is unreached.
src/vllm/model_executor/models/minimax_music3_speech.cpp:638callsMusic3SelectDepthArm. Deleting that call leavestest_minimax_music3_ar37/37 · 640/640 andtest_minimax_music3_speech9/9 · 223/223 green (.agents/specs/minimax-music3.md§19.5 carries the mutation and the binary hashes). The rule it calls is gated on both sides of its condition; what is owed is the call.Why no existing gate sees it. Two reasons, and the second is the one that matters.
--speech-device 1is refused by name on a CPU-only build before a queue exists, so no CI runner can take the branch.ar.depth_stagingis emitted atsrc/vllm/model_executor/models/minimax_music3_llm.cpp:582and no test reads it. §19.6's "device path TAKEN" leg ridestest_minimax_music3_ar, whose observable isMusic3DepthDeviceForwardCount()— which §19.5 itself records as unreachable from production, its only readers being the tests written for it (tests/vllm/models/test_minimax_music3_ar.cpp:1325,1351,1583,1589,1753,1758). A counter no production run reads measures a class, not a capability.It is closable by exactly #1821's method, which is why this is a scoped issue and not a research one. A
tests/parity/test_minimax_music3_depth_arm_real.cppentering throughinclude/vllm.h(vllm_speech_engine_loadwithdevice = 1, thenvllm_synthesize), labelledgpu;checkpoint;music3, exiting 77 without a device or checkpoint, assertingar.depth_stagingcalls == 1from the engine's ownprofile::Reportwith the host bucket absent. #1821's run already firedar.depth_stagingonce onthor:gpu0without asserting it, so the instrument is known to be live on the real path. The acceptance criterion is the same one line: with the engine's call toMusic3SelectDepthArmdeleted, the gate reds.Owned by row
MUSIC3-DEPTH-DEVICE(lane #672). Recorded under## Owedin.agents/specs/minimax-music3.md§19.7, which this issue replaces #1131 as the tracker for.