[bong-water-water-bong]

Carries the fork's current tip — three commits past the pin main holds today (1d32c22609d6, #170), in one PR as asked: #171, #172, #173.

• third_party/llama.cpp: 1d32c22609d6 → d8bd0ff6052a3cc15b5f83d8d0625aac5b644404
  • weekly: 1bit OS image builds under cron; the guard reports real failures #173 hrx: refuse to run when the residency budget is below the working set (engine hrx: with #169's residency budget on, V4.1 loads on HRX0 then SIGSEGVs in CPU SET_ROWS; with the budget off it is RESOURCE_EXHAUSTED at materialize weight 3 (follow-up to #669) #788)
  • docs: census figures of 2026-09-27 (94.88% mapped, 64.33% checked) #172 hrx: make the host-weight residency budget mean the budget (engine hrx: with #169's residency budget on, V4.1 loads on HRX0 then SIGSEGVs in CPU SET_ROWS; with the budget off it is RESOURCE_EXHAUSTED at materialize weight 3 (follow-up to #669) #788)
  • Census 2026-09-27 #171 kv share: an optional guard page after an imported region, so a tile that writes past it faults instead of corrupting
• third_party/hrx-system: unchanged at the held 82804197bcbe (config/hrx-system-hold.config) — the hold is not lifted here.
• registry/architectures.json: regenerated against the new gitlink — 294 hrx / 7 npu, 0 added, 0 removed, --check-pins → "records the committed pins". (The tool reads HEAD, so it runs after the gitlink commit.)

Pre-checked on the new tip: the GGML_HRX_HIP_ADDON_DIR hook is present (6 refs), tests/test-backend-ops.cpp still carries the zero-cases guard, tools/mtmd/models/models.h exists.

Why the gate matters more than usual for this one: #171 changes mappings (a guard page after an imported KV-share region) and #172/#173 change when the engine refuses to run — residency budget versus working set. Both are visible to the hardware gate rather than only to unit tests, so the hw_gate on this branch (triggered by the bump-hrx/* name) is the acceptance: build → 1959/1959 ops → fa-decode 6/6 → kat ×7 → needle ×3 → placement ×7 → kld ×5. The five KLD bases are in the gate store and verified against manifest.json; this bump does not touch them.

Diff vs main is two files (pin + registry) — the gate-hygiene commit from #786 is in main now, so a bump no longer needs to carry it.

Opened, not merged: it lands on the per-bump word.

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 170 - Partially compliant Compliant requirements: Bumped third_party/llama.cpp to the fix for issue HRX: decode-split multipass path intermittently faults (HSA_STATUS_ERROR_MEMORY_FAULT) above capacity 2048 #123 Updated registry/architectures.json to reflect the new commit hash Non-compliant requirements: Documentation in docs/hrx.md was not included in this PR Requires further human verification: Verification that the fix for HSA_STATUS_ERROR_MEMORY_FAULT is correctly applied and tested on hardware 171 - Partially compliant Compliant requirements: Updated census figures in registry/architectures.json Non-compliant requirements: README, PORTING, and docs/registry.md were not updated with the new census figures Requires further human verification: Confirmation that the census data is correctly reflected in all relevant documentation 172 - Partially compliant Compliant requirements: Updated census figures in registry/architectures.json Non-compliant requirements: README, PORTING, and docs/registry.md were not updated with the new census figures Requires further human verification: Confirmation that the census data is correctly reflected in all relevant documentation

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Subproject Commit Update The PR updates the third_party/llama.cpp subproject to a new commit hash. While this is necessary for incorporating the upstream fixes, it's important to verify that the new commit does not introduce any regressions or unexpected behavior in the HRX backend, especially in relation to memory management and MoE router functionality. Subproject commit d8bd0ff6052a3cc15b5f83d8d0625aac5b644404 Architecture Registry Update The registry file has been updated to reflect the new commit hash for llama.cpp (hrx). This change is necessary to ensure that the correct version of the library is being used. However, it's important to confirm that this update aligns with the intended behavior and does not inadvertently affect other parts of the system. "llama.cpp (hrx)": "d8bd0ff6052a3cc15b5f83d8d0625aac5b644404"

[bong-water-water-bong]

hw_gate is GREEN — run 41420550390, ## HRX hardware gate: PASS, 1065 s, 0 amdgpu ring timeouts, Tctl 73 °C at end, on the PR's merge commit 7e6929e (llama.cpp d8bd0ff6052a, hrx-system 82804197bcbe).

row	result
build	PASS 5m18s (CI)
ops	PASS — 1959/1959
fa-decode-q8	PASS 6/6
kat × 7	PASS (9 cases each; 3 and 5 warnings on the two 0.6b runs are the documented ones)
needle × 3	PASS (7341 at 7082 and 7403 tokens)
placement × 7	PASS (all on HRX, largest graph 6449 nodes)
kld × 5	PASS — 0.6b mean 0.000001 / p99 0.00004 / top 100.00%; zaya1-8b, 27b, coder-30b-a3b, 4b-iq4xs −0.000000 / p99 ≤0.00004 / 100.00%

Everything else on the PR is green too — build, CodeQL ×4, both pin guards, and pr_agent passes this time (34 s).

Two things worth noting about this bump specifically, since they are what the gate was for:

• weekly: 1bit OS image builds under cron; the guard reports real failures #173's residency refusal never fired — all seven models loaded and served, so on this box the budget covers each model's working set at the gate's context sizes; the refusal is a floor that these runs stay above, not something that trips in normal service.
• Census 2026-09-27 #171's guard page did not disturb mapping or placement: every placement row is all-on-HRX and no ring timeouts appeared, so the extra page after an imported KV-share region is benign for the paths the gate exercises.

Ready for the per-bump word. Nothing of mine is running; the KLD store was untouched throughout.

[bong-water-water-bong]

VERDICT: READY

At c3c50039 exactly, and its own gate ran on this pin — that is the part worth checking, since a PASS does not transfer
across a re-cut:

hw-gate check at 7e6929e: llama.cpp d8bd0ff6052a, hrx-system 82804197bcbe
## HRX hardware gate: PASS
| build | PASS |  7 s |  | ops | PASS | 124 s | 1959/1959 (MUL_MAT, MUL_MAT_ID, FLASH_ATTN_EXT, GATED_DELTA_NET,
  SSM_CONV, GET_ROWS, ROPE, RMS_NORM, SOFT_MAX) | fa-decode-q8 | PASS | 6/6 |
| kld ×5 | PASS | mean KLD ≤ 0.000001, p99 ≤ 0.00004, same top 100.00% |

So the gate tested d8bd0ff6052a, not the pin it replaces. Content is the two lines it should be: third_party/llama.cpp
1d32c22609d6 → d8bd0ff6052a and registry/architectures.json's llama.cpp (hrx) updated to the same sha — I checked the
two values are equal and that the diff contains nothing else, so hrx-system is untouched (82804197bcbe, the hold file's
value) and no source file rides along.

The pin is the fork tip: d8bd0ff605 = #173 (refuse to run when the residency budget is below the working set) on top of #172
(the capacity-as-budget and ref-shift fixes) on top of #171 (the KV-share guard page) on the TOP_K work — so this bump brings the
whole engine-#788 mitigation set, which is the reason it matters.

I did not update-branch it, as instructed. tprs-ready applied; this is the head to merge on.