The full-refinement extension used the same archived CPU/GPU executables and input SHA-256 values as the GPU4 measurements. The six additional refinements ran sequentially on the same A100 node. A user-owned tmpfs hosted input and output. GPU cases used private CUDA MPS as before.
The ABI-compatible legacy TBB required by the archived public CPU binary was rebuilt from the upstream dependency named by RELION 5.0.1 cmake/BuildTBB.cmake, https://github.com/oneapi-src/oneTBB/archive/v2020.3.tar.gz. Source archive MD5 ea8fa4332f4bad10a75a361cba025380 matched the value in RELION. Apply the bundled RELION cmake/TBB-GCC13.patch; run make -j8 tbb tbb_build_prefix=tbb2020U3; use build/tbb2020U3_release/libtbb.so.2. SHA-256 of the rebuilt library: d28a4f27c42492278a98df7c7237da71f08b8f0895b392e9acd422f904f00c27.

The first 32-core attempt reached iteration 22, but its runtime directory expired before the final maps and timer record were archived. Its duration was excluded. recover32.sh repeats the archived invocation from the original input, with a backup copy after the timed process exits. The 32-core table entry is from the repeat with verified final output.
