cudart: stub Runtime exports required by conda libtorch_cuda

[NickyHeC]

Unblocks #602 conda e2e.

Why

After #602 overmounts shim over libcudart.so.12*, libtorch_cuda.so (conda torch 2.4 / cu12) fails at dlopen on versioned undefs — first hit was cudaDeviceSetLimit@libcudart.so.12. Real cudart is gone, so LD_PRELOAD of the same incomplete shim no longer fallthrough-fixes missing symbols.

Method

objdump -T libtorch_cuda.so  # UND cuda*
nm -D libcudart-shim.so      # T exports
comm -23

→ 8 missing vs current shim. Stub only those.

Exports added

Symbol	Behavior
cudaDeviceSetLimit	success no-op
cudaDeviceGetLimit	write 8 MiB (mirrors cuCtxGetLimit)
cudaGraphAdd{Kernel,Host,EventRecord,EventWait}Node	NotSupported (801); stream-capture remains the forwarded path
cudaGraphRetainUserObject	success no-op
cudaUserObjectCreate	sentinel handle
cudaStreamUpdateCaptureDependencies	success no-op

Proof — Lambda A10, #602 agent + this shim, no LD_PRELOAD

Image: pytorch/pytorch:2.4.0-cuda12.4-cudnn9-runtime

cuda: True
matmul_ok torch.Size([4, 4])
backward ok   # transformers tiny-random-LlamaForCausalLM

pip portal-cuda control unchanged (cuda: True) — that link set didn't require these symbols at import.

[ghost]

Ready to review this PR? Stage has broken it down into 2 individual chapters for you:

	Title
1	Stub device limit APIs for PyTorch compatibility
2	Add stubs for CUDA graph and stream capture

_{Chapters generated by Stage for commit 8d7e0d6 on Jul 15, 2026 7:10am UTC.}

[NickyHeC]

Lambda retest — PASS

On A10 / Ubuntu 22.04 with #602 agent + this shim in packaged rootfs, image pytorch/pytorch:2.4.0-cuda12.4-cudnn9-runtime, no LD_PRELOAD:

cuda: True
matmul_ok torch.Size([4, 4])
backward ok   # HF tiny-random-LlamaForCausalLM + transformers

Audit method: objdump -T libtorch_cuda.so UND cuda* vs nm -D shim exports → 8 missing symbols stubbed across two commits on this branch (79d9ed52, 8d7e0d66).

[NickyHeC]

Evidence (see description).

Audit: 8 UND cuda* symbols in conda libtorch_cuda.so missing from shim → stubbed.

#602 agent + this shim, pytorch/pytorch:2.4.0-cuda12.4-cudnn9-runtime, no LD_PRELOAD:
cuda: True → matmul → tiny Llama backward ok.

[NickyHeC]

Companion to #602 (stubs the 8 conda libtorch_cuda.so Runtime exports missing from the shim). Combined #602 + #638 is the validated e2e config (cuda: True → matmul → tiny-Llama backward ok, no LD_PRELOAD). Marking ready for review; best landed together with #602.