fix: keep hrx-system at 98d05d94 (our loom-jit needs its AMDGPU runtime-globals API)

[bong-water-water-bong]

Current main does not compile with HRX. #353/#357 moved third_party/hrx-system to 4ba76c18eafe (our fork's 1bit/main, which merged AMD's c0b135a778cc). That merge carries AMD's loom commit 32d0c76a "[Loom] Make AMDGPU runtime globals compiler-owned", which deletes loomc_amdgpu_runtime_global_flags_t and loomc_amdgpu_emit_options_t from loom/binding/c/include/loomc/target/amdgpu/emit.h.

Our ggml/src/ggml-hrx/loom-jit.cpp still uses them, so the pinned pair fails:

loom-jit.cpp:70:5:  error: unknown type name 'loomc_amdgpu_runtime_global_flags_t'
loom-jit.cpp:557:59: error: use of undeclared identifier 'LOOMC_AMDGPU_RUNTIME_GLOBAL_FEEDBACK_CONFIG'
loom-jit.cpp:994:5:  error: unknown type name 'loomc_amdgpu_emit_options_t'
10 errors generated.

GitHub-hosted CI builds the pin without HRX, so it cannot catch this (see bump-hrx.yml); a from-scratch HRX build on strixhalo is what surfaced it.

Fix: restore third_party/hrx-system to 98d05d94a9f9e405683f8aeafc037f5ace0abe4a (our fork's previous revision, which exports the API at loom/binding/c/include/loomc/target/amdgpu/emit.h:37). Verified: recompiling loom-jit.cpp against 98d05d94 succeeds (rc=0). This is the constraint already recorded in OWNER-ACTIONS.md §"engine#336": "leaves hrx-system at 98d05d94 — AMD's hrx-system does not export loomc_amdgpu_runtime_global_flags_t, which our loom-jit path needs."

Pin rollback on purpose. llama.cpp is unchanged at e44c9d01a4d5.

Follow-up (forward direction): port loom-jit.cpp to the compiler-owned API and move hrx-system forward again — tracked separately so the release pin can build now.

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 353 - Partially compliant Compliant requirements: Move third_party/llama.cpp to the correct commit Move third_party/hrx-system to the correct commit Resolve conflicts as described No changes to .github/workflows CI builds without HRX Non-compliant requirements: Validation on Strix Halo with test-backend-ops -b HRX0 not explicitly shown in diff but required by ticket Requires further human verification: Validation on Strix Halo with test-backend-ops -b HRX0 needs to be confirmed by human 357 - Partially compliant Compliant requirements: Both submodules moved to AMD-pinned pair Our own commits applied Conflicts resolved as described No changes to .github/workflows CI builds without HRX Non-compliant requirements: Validation on Strix Halo with test-backend-ops -b HRX0 not explicitly shown in diff but required by ticket Requires further human verification: Validation on Strix Halo with test-backend-ops -b HRX0 needs to be confirmed by human 336 - Partially compliant Compliant requirements: Pin third_party/llama.cpp to correct commit Pin third_party/hrx-system to correct commit to retain needed API Verified with test-backend-ops -b HRX0 Verified recompilation of loom-jit.cpp succeeds Non-compliant requirements: Validation on Strix Halo with test-backend-ops -b HRX0 not explicitly shown in diff but required by ticket Requires further human verification: Validation on Strix Halo with test-backend-ops -b HRX0 needs to be confirmed by human

⏱️ Estimated effort to review: 2 🔵🔵⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review API Compatibility Issue The PR reverts third_party/hrx-system to commit 98d05d94 to maintain compatibility with loom-jit.cpp, which depends on loomc_amdgpu_runtime_global_flags_t and loomc_amdgpu_emit_options_t that were removed in the newer commit 4ba76c18eafe. This is a necessary compatibility fix to prevent compilation errors when building with HRX enabled. Subproject commit 98d05d94a9f9e405683f8aeafc037f5ace0abe4a