CDSP transport evidence ledger
This ledger distinguishes code-execution, transport, and performance claims. It does not select a product transport.
| Evidence | Established | Not established |
|---|---|---|
r0sub0-cross-soc-20260924.md |
The same emitted R0Sub0 kernel ran on SM8550 and SM8635. On SM8635 its DSP-tick median was 2.286–2.302x faster than the LLVM comparison kernel; both were invoked through libcdsprpc.so. |
A FastRPC bypass, a complete model, or a whole-path speech speedup. |
direct-fastrpc-native-open-20260925.md |
A direct libc open of an ADSP-named node returned EACCES in both tested app sandboxes. |
CDSP access through that node, another node, or a lower transport. |
| Owner-reported cross-device smoke test | The owner reports a Razr+ per-job FastRPC-bypass speedup and an S23 failure. Related diagnostic queue scripts and current results were recovered from both phones; see recovered-dspqueue-diagnostics-20260926.md. |
Whether the reported S23 failure was an earlier revision or a different test; the failing stage is not in the inspected current results. |
| Pinned Qualcomm DSPQueue source | Defines a QNN-independent shared-memory queue layout, a FastRPC queue bootstrap, and driver-signaling capability checks. See dspqueue-upstream-audit-20260926.md. |
Use of this mechanism in the reported phone tests, Queue Monitor availability on either phone, and a universal no-ioctl-per-job path. |
s23-cdsp-node-inventory-20260926.md |
In a read-only S23 shell check, only ADSP-named RPC nodes appeared under the queried paths; the kernel base differs from the external report's LineageOS candidate. | App access, a queue route, the prior S23 failure stage, or a Razr+ comparison. |
cross-device-cdsp-inventory-20260926.md |
A same-session read-only check found matching queried RPC nodes, modes, owners, and labels on both devices; neither installed kernel base matches the external candidate revisions. | Equal firmware capabilities, app access, or the cause of the reported queue-test split. |
recovered-dspqueue-diagnostics-20260926.md |
Both diagnostic echo results pass. The Razr+ queue median was 1.94x faster than its synchronous-invoke median; the S23 queue median was 1.21x faster than its baseline. | Identical echo scripts, zero library-internal FastRPC activity, glitch-free audio on both phones, or Kokoro synthesis. |
Next evidence action: preserve a reviewed source-defined queue contract in this repository, then test the same owned worker and queue protocol on both devices. Keep the historical library-mediated echo and audio probes separate from that product gate. The earlier S23 failure stage remains unlocated. Do not infer that FastRPC is mandatory or that a kernel modification is needed.