#2457 0.1.44 on HIP/gfx1030 (RX 6950 XT, Windows): IQ3_S output turns to garbage, Q2_0 fine, 0.1.43.1 fine

Description

Hi, and thanks for Strata, it's been running really well on my card.

After updating to 0.1.44, the IQ3_S pack of Qwen3.8-Flash-Next started giving me nonsense on my RX 6950 XT. A reply would start out roughly on topic and then drift into stray tool-call tags, random Chinese, or loop on the same phrase until it hit the token limit. Nothing in the log looks wrong. The engine starts, the verify windows pass, it just says the wrong things.

What I checked:

- Same machine, same 0.1.44 engine, the Q2_0 pack answers normally. So it isn't the whole engine on this GPU.
- I rolled back with `START-HERE.bat --rollback-engine` to 0.1.43.1. Same IQ3_S files, same config, and the output is correct again.
- On a small eval set I use (12 domain questions), IQ3_S scored 11/12 on 0.1.43.1 and 0/12 on 0.1.44.

Something in the logs that might help: on 0.1.44 the replies were much longer and drafts were accepted a lot, for example `196 generated ... drafts accepted 99 of 146`. On 0.1.43.1, similar prompts give `36 generated ... drafts accepted 23 of 28` and `40 generated ... 26 of 33`. That fits the looping I was seeing. Startup looks the same on both versions apart from the version line (expert cache 3437 slots, R4 hit path ON, MTP draft layer loaded).

I haven't tried turning off MTP, speculative decoding or the expert cache on 0.1.44 to narrow it down. I still have 0.1.44 in `engine\.previous`, so I'm happy to run whatever combination would help.

**Setup**

- Windows 11 Enterprise, build 26300
- AMD Radeon RX 6950 XT (gfx1030), 16 GB, driver 32.0.21045.5002, 64 GB RAM
- Prebuilt HIP engine. Both builds report rocm `10.2.0a20260930` with gfx1030 in `archs`
- Model: Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S (2 GGUF parts) with the native pack

<details><summary>Engine args (from the config)</summary>

```
--pack C:\llm\strata-data\packs\iq3_s
--native C:\llm\models\qwen38-IQ3_S\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf
--ple-gguf C:\llm\models\qwen38-IQ3_S\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00002-of-00002.gguf
--expert-profile C:\llm\strata\data\expert-profile.bin
--expert-cache auto --prefill auto
--spec 4 --spec-min-p 0.5
--mtp C:\llm\strata-data\mtp\rt
--max-context 32768 --kv int8
backend hip, gpu 1
```
</details>