[AaronStGeorge]

The recorded HRX minimum predates dependencies already used by this backend: the loom-benchmark build target and check.expect.approximate in the kernel corpus. Raise hrx-minimum-commit.txt from 771711d49befd397f10c3c6deb66f1c5c89e5042 to d03575e84f602d3e39b87895a30fae214f308558.

d03575e84f first adds the required typed expectation operation (merged by 7c0f021226, HRX ggml-org#2418). It contains both the old floor and the earlier execution-tool rename bdf3e0ceee in its ancestry. Its immediate parent, d270efd38c, rejects the exact expectation statement from gather_add_f32.loom; the selected revision parses it successfully. This selects the earliest demonstrated dependency revision, before the previously tested 82214dfb02.

Validation against llama.cpp 2120f7434 plus this one-line change, using HRX at the exact new floor:

• Fresh Release build with ROCm SDK 10.1.0a20260822-1100, ROCm Clang, and compiler caching disabled. Built HRX/Loom, the binary kernel corpus, all four HRX suites, llama-bench, and llama-cli through GGML_HRX=ON and HRX_SOURCE_DIR.
• Revision checker accepts the exact floor and rejects both the old floor and the immediate parent. CMake accepts the exact floor and rejects the parent during configuration.
• ROCR_VISIBLE_DEVICES=GPU-fc2143878bea4591 ctest --test-dir <build> -L '^hrx$' --no-tests=error --parallel 2 --verbose: all four suites pass on gfx1100 RX 7900 XTX. Buffer 5/5, backend 106/106, Loom JIT 6/6, operations 334/334; 451 cases, zero failures.
• Shared Llama-3.2-1B-Instruct-UD-Q4_K_XL.gguf: llama-bench -ngl 99 -dev HRX0 -p 128 -n 32 -r 1 completes prefill/decode. Greedy llama-cli generation on HRX0 answers the arithmetic prompt with "Two plus two equals four."
• git diff --check passes; only the minimum-commit file changes.

As documented in the backend README, this minimum does not guarantee compatibility with future breaking HRX revisions.

AI usage: Codex audited the history, made the change, ran validation, and prepared this PR under explicit user direction.