#1715 [Strix Halo, 128 GB unified memory] setup's --expert-cache auto (no --resident-budget-gib) causes a global OOM with UD-IQ4_XS

Descripción

**Setup:** Strix Halo laptop (HP ZBook Ultra G1a, Ryzen AI Max+ 395, 128 GB, gfx1151), Ubuntu 26.04, kernel 7.0.0-38, v0.1.41 source build per `docs/STRIX_HALO.md` (ROCm 7.14.1 tarball), Unsloth UD-IQ4_XS, host `ttm.pages_limit=29360128` (112 GiB GTT) and `amd_iommu=off`, desktop session running.

**What happened:** an engine run with the arguments setup writes for this model — `--expert-cache auto` and no `--resident-budget-gib` (otherwise `--prefill 16384 --spec 4 --spec-min-p 0.5 --mtp ... --kv int8 --lookup-chain 3 --mtp-q4 all`, plus the `docs/STRIX_HALO.md` §5 fast config) — triggered a **global OOM**. The kernel killed the engine and processes in the desktop session:

```
kworker/0:2 invoked oom-killer: gfp_mask=0x2cc2(GFP_KERNEL|__GFP_HIGHMEM|__GFP_NOWARN), order=0, oom_score_adj=0
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),cpuset=/,mems_allowed=0,global_oom,task_memcg=/user.slice/...,task=strata,pid=21418,uid=1000
Out of memory: Killed process 21418 (strata) total-vm:132671880kB, anon-rss:53770192kB, file-rss:8kB, shmem-rss:207500kB, UID:1000 pgtables:117032kB oom_score_adj:200
```

**Expected:** `auto` leaves enough room on a unified-memory APU. `docs/STRIX_HALO.md` says `device_free_bytes()` counts `MemAvailable` less 6 GiB; here that still overcommitted — possibly because the expert cache and the resident/host copies both draw on the same 128 GB, or because the larger GTT pool makes "free VRAM" look bigger than what RAM can back.

**Workaround:** explicit `--expert-cache 24576 --resident-budget-gib 55` is stable (≥45 GiB `MemAvailable` throughout agentic and 128K-context runs; all 24,576 experts resident).

Benchmark results from the same machine are in a separate PR (community report). Write-up: https://github.com/nsbradley88/halo-tuning