#3599 [Engine] strata.exe exits silently during long-prompt prefill after an over-context request in the same session (gfx1031, Windows)

Descrição

### Area

[Engine] (speed, kernels, models, the server and API)

### Desktop app, Terminal or source?

Source (START-HERE.bat, setup.sh or a build)

### What happened

Within a single server session, `strata.exe` **exited silently** while reading a long prompt - no error on stdout, no error in `strata-<model>.log`, the process is just gone (Windows Error Reporting has a bucket for strata.exe).

Observed sequence (benchmark harness against the local server):

1. Three 4,196-token requests processed normally.
2. One deliberately oversized 53K-token request rejected with 400 ("request exceeds the context window ... requests are never truncated").
3. Next request: 16,685 tokens. Server stdout shows the progress lines `reading the prompt: 6,144 of 16,685 tokens` ... `12,288 of 16,685 tokens` - then nothing. The process died between the 12,288-token progress line and completion, ~48 s into prefill.

Seen twice total (the second time during normal GUI use, also after earlier long requests in the same session). A cold server handling the same 16,680-token request as its first request works fine (prefill 314.7 tok/s). My working theory is leftover state from the rejected oversized request (it had already started reading 53K tokens into the same session before the 400), but that is a guess - writing only what I saw here.

### Smallest steps or command that reproduce it

Could not make it happen on demand. Closest sequence (single server session, port 8081):

1. Start Qwen3.8-Flash-Next IQ3_XXS (32K context at the time).
2. POST three ~4K-token chat requests (OK).
3. POST a request that exceeds the context window (rejected 400 by the server).
4. POST a ~16.7K-token request -> engine process exits silently mid-prefill (both times after step 3 was present in the session history).

A fresh server run with only step 4 does not crash. Logs and bench-output JSON available on request.

### Reproduction status

Reproduced locally (intermittent: two occurrences, both after a prior oversized/long request in the same session; not reproducible on demand).

### Strata version

Strata 0.1.45, engine 0.1.45 - self-built from the v0.1.45 source with `tools\hip\build_windows.bat` (gfx1031 only, ROCm 10.2.0a20260930, llama.cpp snapshot 3cf03257)

### Graphics card, memory and system

AMD Radeon RX 6700 XT 12 GB (gfx1031) · 64 GB RAM · Windows 10 Pro 22H2 (19045)

### Model, size and settings

Qwen3.8-Flash-Next IQ3_XXS · 32K context at the time (also happens on 64K config after a long session) · `--kv int8` · MTP draft layer on · pictures off

### Logs

Engine log ends cleanly on the last successful request - no trace of the crash; the only evidence is the frozen `reading the prompt: 12,288 of 16,685 tokens` progress line on server stdout and the WER bucket. Full `strata-iq3_xxs.log` and server stdout captured at: `D:\strata-6700xt\serve-bench.log` (can paste or attach on request).