#3463 [Engine] RX 7900 XTX (gfx1100) machine restarts under Strata serving (follow-up to #872)

Descripción

Follow-up to #872.

## Hardware
- Dell Precision 3660, i9-12900K, 64 GB DDR5 (4×16 GB, 4800 MT/s modules; runs at 4000 MT/s with four DIMMs)
- AMD Radeon RX 7900 XTX 24 GB (gfx1100), compute-only; display on the Intel iGPU
- Dell 1000 W Platinum PSU

## Software
- Ubuntu 26.04 LTS, kernel 7.0.0-38-generic, amdgpu kernel driver 3.64.0
- Strata engine 0.1.45 and 0.1.45.3 (HIP, gfx1100, built from source); ROCm userspace from the Strata venv

## Symptom
While Strata is serving, the machine self-restarts. No shutdown, panic, OOM, MCE, thermal or amdgpu message is logged; the journal ends mid-activity and the engine log ends mid-request. The machine returns after a cold boot. 18 restarts on Linux so far, 2.4–88 min after the load starts.

## Test prompt
The prompt loops: read 80 lines from a source file, append a short function (≤15 lines), run python3, and repeat.

## Configurations tested
- Qwen3.8-Flash-Next IQ3_S, 262144 context: restarts.
- Qwen3.8-Flash-Next IQ3_S, 131072 context: restarts.
- IQ3_S low-RAM, 262144 context (`--resident-budget-gib 42`, `--pcie-frac 0`, `--vram-reserve-mib 832`, `HSA_USERPTR_FOR_PAGED_MEM=0`): 36.37 GiB pinned, 7,317 experts in VRAM, file tier idle; 45–46 GiB RAM used / 14–15 GiB available. Survived 73.8 min, then restarted after 12 cycles of the test prompt in the next run.
- Qwen3.8-Flash-Next IQ3_XXS, 131072 context: 39.97 GiB arena, 8,914 experts / 14.43 GiB in VRAM; 47 GiB RAM used / 13 GiB available. Restarted after 8.5 min, and again 2.5 min after a fresh engine start.
- `--pcie-frac 0` (IQ3_XXS, 131072): decode 93–110 tok/s versus ~80 at baseline at the same context; prefill unchanged at ~1,140 tok/s. Restarted at 23m40s.
- BIOS PCIe link speed Gen 3: restarted after ~50 min.
- BIOS ASPM disabled: restarted after 10m14s. The BIOS setting did not clear ASPM L1 on the links (they still reported `ASPM L1 Enabled`); ASPM was also disabled with `setpci`, and the restarts continued.
- MemTest86+: 8 full passes, 0 errors.

With the test prompt, the machine also restarted at 22m56s, 5m56s, and 12m32s in the other attempts.