#3425 [Engine] Recommended IQ3_XXS on 32 GB RAM + 16 GB VRAM generates at 0.5–7 t/s

Description

### Problem
The size Strata recommends for 48 GB RAM + VRAM (IQ3_XXS, 47.0 GB) fills system RAM on a 32 GB RAM + 16 GB VRAM PC. Generation runs at 5–7 t/s and drops to 0.5 t/s as the context grows, against the published 62 t/s. The GPU stays almost idle. A simple question in the built-in Strata chat did not finish in 20 minutes.

### Environment
- Strata version / commit: <version>
- OS: Windows 11 23H2 (build 22631)
- GPU: NVIDIA RTX 5080 16 GB, driver 617.14, CUDA 13.4, PCIe Gen3 x16
- CPU: AMD Ryzen 7 5700X (8C/16T)
- RAM: 32 GB DDR4 (XMP: <on/off>)
- Storage: W:\Strata\Strata-data on <NVMe / SATA SSD>
- Model: Qwen3.8-Flash-Next IQ3_XXS (GSQ-RCO)
- Settings: context 64K, resource mode "Leave 20% for my work", image reading on
- Client: the built-in Strata GUI chat with its agent tools (workspace "Strata")

### Reproduce
1. On the PC above, pick the model recommended for 48 GB (IQ3_XXS) and keep the defaults listed above.
2. The Strata UI shows "43.9 GB RAM + VRAM" and "About 10 GB will be read from SSD, so answers may be slower".
3. In a new Strata chat, attach a screenshot and ask: "how do I add the Strata model to OpenCode" (asked in Russian).

### Expected
An answer within a few minutes, with generation close to the published 62 t/s for IQ3_XXS. Otherwise, a recommended size that leaves enough free RAM for a normal desktop session.

### Actual
- The model thinks for 2m 41s, then calls tools: the glob `**/opencode*` times out ("Glob timed out (huge home dir)"), and two pwsh commands run.
- After that the status stays at "Preparing answer..." for 19m 10s with only 869 tokens generated, and no answer is ever given.
- Request log from the Strata dashboard for the same session:

~~~text
time      prompt   reused   output   gen t/s   duration
14:21:38      53        0        1       7.6     16.9 s
14:25:12   7,068        0      897       5.1    5m 55s
14:31:40   8,010    7,964      142       5.8     34.4 s
14:39:49   8,317    8,152      144       5.9     45.1 s
14:40:43  19,139    8,462        2       0.5    3m 42s
~~~
- Prompt processing: 35–50 t/s, while the published figure is 1,750 t/s.

### Evidence
- System RAM: 25.9 / 32 GB after the first requests, 31.0 / 32 GB at 19K context.
- VRAM: 14.1–14.2 / 16 GB; experts in VRAM: 3,405 (5.6 GB).
- GPU load during generation: 6–8 %, power draw 47–58 W out of 360 W.
- PCIe traffic to the GPU: 18–48 MB/s; disk read: 4–6 MB/s.
- The "VRAM hit rate" column is empty ("—") for every request.
- Log: strata-<model>.log attached (sanitized). Screenshots of the chat and the dashboard attached.

Same PC with plain llama.cpp and a model that fits fully in VRAM, used as a coding agent in OpenCode:
~~~bat
llama-server.exe -hf unsloth/Qwen3.8-27B-GGUF:UD-IQ3_XXS ^
  -ngl 99 -c 65536 -fa on -np 1 --jinja --cache-ram 0 ^
  -ctk q8_0 -ctv q8_0 ^
  --spec-type draft-mtp --spec-draft-n-max 2
~~~
Result: 30–70 t/s generation and stable tool calls; a question like this one is answered in a couple of minutes. So the hardware can run a coding model quickly; the slowdown comes from the recommended Strata size not fitting in free memory.

Suggestions:
- Subtract the memory already in use (Windows, browser and IDE take 6–10 GB) when recommending a size, or warn when less than about 8 GB would stay free.
- Show the expert cache hit rate in the UI, so users can tell whether VRAM, RAM or SSD is the bottleneck.

### Reproduction status
Reproduced by me on the hardware above, in one session with several requests. Not yet retested with "Full performance" mode, XMP on, or with other applications closed.


### Test results (update)

Short prompt without tools, new chat in the Strata GUI, same IQ3_XXS model:

> Write a C++17 function that parses a CSV line with quoted fields. No explanation.

~~~text
request                               output   duration   average
CSV function (C++17)                     454    3m 4s     ~2.5 t/s
"how to add the model to OpenCode"     1,183   19m 31s    ~1.0 t/s
~~~

- Settings: context 64K, "Read pictures" on.
- The Strata UI warns: "About 10 GB will be read from the SSD, so answers can be a little slower."
- The dashboard shows Speed as "Unavailable" for these requests, so I could only compute the averages (output / duration). The first figure may include model load time.
- After the requests the model was unloaded ("No model loaded"; RAM 7.0 / 32 GB, VRAM 1.2 / 16 GB).

Storage note: all model files are on an external SSD connected over USB Type-C. llama.cpp loads Qwen3.8-27B from the same drive and still generates at 30–70 t/s, because that model fits fully in VRAM and the drive is read only at load time. With Strata the recommended size relies on reading about 10 GB from this drive during generation, and the UI gives no warning that external or USB storage is unsuitable for that.