[TATP-233]

Summary

• Add an explicit event_wait="stream" | "synchronize" mode to CudaRingTransport.
• Select synchronize on HIP/ROCm and keep stream waits on NVIDIA CUDA.
• Document the ROCm compatibility path as a performance-impacting fallback: unfixed ROCm CLR 7.2 emulated IPC event streamWaitEvent omits the 32-entry signal-ring modulo, so the 33rd reuse addresses signal[32] and fails with hipErrorInvalidValue.
• Add unit coverage for runtime selection and host-event synchronization plus a real HIP slow test that crosses the 33rd event reuse.

This partially addresses #2530; the upstream runtime and benchmark-provenance items remain tracked there.

Validation

uv run --no-sync pytest -q tests/rl/runtime/ipc/test_inference_ring.py
# 19 passed, 4 deselected

uv run --no-sync mypy src/unilab
# passed

uv run --no-sync pyright
# 0 errors

make check
# passed

make test-all
# 2833 passed, 61 skipped, 375 deselected, 18 warnings
# benchmark import smoke: 35/35 modules, 36/36 scripts

Remote MI300X / ROCm 7.2 / Genesis 1.3.3:

ROCR_VISIBLE_DEVICES=0 uv run --no-sync pytest -q -m slow \
  tests/rl/runtime/ipc/test_inference_ring.py::test_hip_inference_ring_survives_rocm_ipc_event_slot_reuse
# 1 passed

Production-shape learner-coupled SAC run:

ROCR_VISIBLE_DEVICES=0 timeout 600 .venv/bin/train \
  --algo sac --task g1_walk_flat --sim genesis \
  algo.num_envs=2048 algo.batch_size=256 \
  algo.replay_buffer_n=4 algo.learning_starts=1 \
  algo.updates_per_step=1 algo.max_iterations=128 \
  algo.save_interval=1000 algo.algo_params.use_compile=false \
  training.no_play=true training.export_onnx=false \
  training.log_dir=/tmp/unilab-rocm-genesis-ipc-fix-2048
# exit 0; completed_iterations=128; total_env_steps=264,192
# tail_env_steps_per_sec=34,690.9; final_env_steps_per_sec=36,000.4
# normal_completion, no dropped replay batches, collector exit 0

Small 64-env run also crossed the former tick-32 failure:

completed_iterations=128
total_env_steps=8,448
tail_env_steps_per_sec=1,257.4
final_cycle_wall_ms=51.74
status=completed