Upgrade the Primus runtime image from rocm/primus:v26.3 to v26.4 across CI Dockerfile/workflow, example launch scripts, benchmark scripts and docs, drop v26.3-only CI workarounds now that they're no longer needed, and fix several bugs found while validating the change end-to-end. JAX/MaxText base image not upgraded this round.

Image bump

• .github/workflows/ci.yaml & docker/Dockerfile: BASE_IMAGE -> v26.4, sync v26.3 comments
• README/examples/docs/benchmark scripts: bump documented image tag to v26.4, release branch to release/v26.4, pip pin to primus==26.4.0
• runs-on stays on the v26.3 self-hosted runner label (no v26.4 runner registered yet)

Workaround cleanup

• Drop the v26.3-only "Purge stale TorchTitan install" step: v26.4 no longer bundles a torchtitan that shadows the submodule.
• Drop the origami (rocm-libraries@223648a) install from Dockerfile and CI: this base image ships no origami, so primus_turbo falls back to heuristic kernel selection for MoE grouped-gemm. Left the pip install one-liner in a comment to add it back if ever needed.
• aiter_deepbind_patches.py self-detects whether the RTLD_DEEPBIND workaround is still needed: skips installing it once transformer_engine.__version__ >= 2.14 (fixed in v26.4), still applies on v26.3 and older. Added unit tests for the gating logic.
• Drop the from-source rocSHMEM rebuild and the v26.3 ROCM_HOME/UCX_HOME/MPI_HOME overrides: v26.4 already ships a working prebuilt rocSHMEM and correct env vars for its pip-installed ROCm SDK layout; the old overrides pointed at a now-nonexistent /opt/rocm and broke the rebuild.

v26.4 pip-installed ROCm SDK fallout

v26.4 switched from a filesystem-installed /opt/rocm to a pip-installed rocm-sdk-devel venv package. Several things assumed the old layout:

• Dockerfile.ainic: ROCM_PATH/MPI_PATH were hardcoded to v26.3 paths; rccl/amd-anp also hardcode /opt/rocm internally, and amdclang++ couldn't find its sibling clang++ (moved under llvm/bin/). Fixed via symlinks instead of patching those repos, plus added pipefail so cmd | tee log steps fail on cmd's exit code (was silently swallowing failures until a much later, unrelated-looking step).
• torch is now a proper pip wheel with a strict Requires-Dist: triton==<exact commit> in its metadata. Our Dockerfile force-reinstalls a different pinned triton commit without updating that metadata, so any later pip install (e.g. the unit-test job's requirements.txt) saw torch's declared triton dependency as unsatisfiable and silently replaced torch with a plain PyPI/CUDA build, breaking torchvision and other ROCm extensions. Fixed by patching torch's metadata to the actually-installed triton version right after building it. Also restored the TRITON_COMMIT build-arg, which had been dead (hardcoded to 09500db9 in the Dockerfile) since Build Triton from source and bump Primus-Turbo  #694.

CI fix

• The unit-test container reuse check compared only the image tag string (docker.io/tasimage/primus:${IMAGE_TAG}), not content. On main, IMAGE_TAG is always latest, so after build-docker repushes under that tag, the check still saw an identical string and reused the stale container instead of pulling the new image. Fixed by pulling first and comparing the resolved image id instead of the tag name.