#1860 [Bug]: HIP engine aborts with SIGABRT (std::thread -> abort) on RX 6800/6900 (gfx1030), 16 GB VRAM + 31 GB RAM low-RAM mode

Description

### What happened

Fresh install on Linux following `docs/AI_SETUP.md` (the agent flow). Setup and first start were fine, then the engine died with `SIGABRT` during normal use and the host got unstable enough that I uninstalled.

Steps:

```
./setup.sh --yes --family coder --model IQ1_M --no-start
./run-coder-iq1_m.sh          # (server on a non-default port, see below)
```

- The server came up: `GET /health` -> `{"status":"ok", ..., "loaded":true, "service":"strata"}`.
- `GET /v1/models` listed `qwen3.8-flash-next-coder-iq1_m`.
- One `POST /v1/chat/completions` (`reasoning_effort: none`) answered correctly at ~6.5 tok/s.
- Later, during use, the engine process aborted and systemd-coredump captured a core.

Expected: a stable server.

Port note (probably unrelated): port 8080 was already taken by an unrelated service on this box, so I reconfigured the install to `--port 8081` and served there.

### Crash data (`coredumpctl info`)

```
PID: 3315559 (strata)
Signal: 6 (ABRT)  si_code: SI_TKILL        # abort()/raise, not a SIGSEGV
Command Line: /home/vpinto/Strata/engine/strata --serve \
  --pack /home/vpinto/Strata-data/packs/coder-iq1_m \
  --native .../Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M-00001-of-00002.gguf \
  --ple-gguf .../Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M-00002-of-00002.gguf \
  --expert-profile /home/vpinto/Strata/data/expert-profile-coder.bin \
  --expert-cache auto --prefill auto --spec 4 --spec-min-p 0.5 \
  --mtp /home/vpinto/Strata-data/mtp/rt --max-context 65536 --kv int8 --resident-experts

Stack trace of thread 3316197:
  #0  libc.so.6 + 0x9a17c
  #1  raise            (libc.so.6)
  #2  abort            (libc.so.6)
  #3  strata + 0xe0c335  _ZNSt6thread11_State_implINS_8_InvokerISt5tupleIJZ4mainEUlvE33_EEEEE6_M_runEv
  #4  libstdc++.so.6 + 0xea859
  #5  libc.so.6 + 0x980a2
```

So `abort()` is reached from a worker `std::thread` whose entry is a `main()` lambda (`main::$_33`) — consistent with a fatal error path or an uncaught exception in a thread.

The core file is still on disk: `/var/lib/systemd/coredump/core.strata.1000.cc9a45f591e04a0b9e3aca3a3e384547.3315559.1791648608000000.zst` (~1.7 GB).

### Strata version, GPU, OS

- Strata: `main` cloned 2026-10-10 (shortly before v0.1.42 was tagged; the version string was not recorded before uninstall — most likely 0.1.41). Engine build `llama.cpp 3cf0325`.
- GPU: AMD Radeon RX 6800 / 6900 series, 16 GB VRAM, `gfx1030`, single card, **also drives the desktop**.
- OS: Omarchy 4.0.4 (Arch), kernel `amdgpu` driver, no system ROCm.
- CPU/RAM: Ryzen 7 5800X (AVX2), 31 GB RAM.
- Backend: HIP. Setup installed ROCm `7.13.0a20260515` (TheRock gfx103X wheels) into `.venv`.
- Setup choices: Coder `IQ1_M`, context `65536`, `--kv int8`, images off, low-RAM mode (GPU holds ~45% of the experts).

### Notes that may matter

- With the model loaded the server reported only **616 MiB of VRAM free**, and the same GPU drives the desktop.
- RAM was at the edge (31 GB total; the engine warns the ~32 GB IQ1_M plus ~0.9 GB KV streaming did not fit, so KV stayed in VRAM).
- Possibly related open issues I found while searching: #937 (`[gfx1030] HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION in gdn_step_commit_kernel ...`), #1103 (`gfx1030 ... ROCm 7.14 gfx103X wheels regression, fixed by 7.13`), #649 (`HIP (gfx1030): intermittent 'verify: timed out at layer N' ...`), #516 (compositor VRAM exhaustion when `--expert-cache auto` fills a card that drives the desktop).

### Engine log

Not available — I removed the install (folder + data) before collecting `strata-coder-iq1_m.log`, so I have only the coredump metadata above and the core file. Happy to run any diagnostic against the core or reinstall a specific version if that helps. I can also attach the core if there is a way to share it.