[BUG] Gemma 4 26B misidentifies environment warnings as crash root cause in WinDbg !analyze output

### Model
gemma-4-26B-A4B-it (Q4_K_M via llama.cpp b8668)

### Behavior
I gave Gemma 4 a WinDbg `!analyze -v` crash dump and asked "Analyze this crash dump and identify the root cause." It consistently reported the extension loading warnings at the top of the output (`LoadLibrary(ext) failed`, `No export analyze found`) as the crash cause, completely missing the actual `*** Bugcheck Analysis ***` section further down (BugCheck 0x7F — stack overflow from 447 recursive calls in StackOverflowDriver).

Ran the same test 10 times with the same input — Gemma 4 missed the real root cause every time. Qwen 3.5 35B and Claude Opus both correctly identified the stack overflow on every attempt.

Adding an explicit instruction to ignore LoadLibrary warnings helped (went from always wrong to mostly right), but that shouldn't be necessary — the other models handle it without hints.

### What I think is happening
The WinDbg output starts with ~15 lines of extension loading messages (normal debugger startup noise), then the actual analysis begins at `*** Bugcheck Analysis ***`. Gemma 4 seems to latch onto the early error-looking text and never gets to the real analysis section.

### Environment
- Inference: llama.cpp b8668
- Quantization: Q4_K_M (GGUF)
- Prompt: "Analyze this crash dump and identify the root cause"
- Tested: 2026-04-08