#2396 0.1.44 refuses GGUFs whose attention.compress_ratios is all zeros (OrcaRouter Qwen3.8-Flash-Next-Uncensored)

本文

Since 0.1.44 the engine refuses the GGUF files of `orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF` (an abliterated Qwen3.8-Flash-Next), which 0.1.43 ran without trouble:

```
qwen4exp.attention.compress_ratios[3] = 0, but full_attention_interval = 4 makes layer 3 full attention
```

**Cause.** These files carry `qwen4exp.attention.compress_ratios` as 48 zeros, one per layer. That looks like a placeholder written by the converter, not a layer split: every entry is 0, including the full-attention layers. 0.1.43 did not read the key and ran the files with the compiled pooling block. 0.1.44's new check in `read_geometry` (`include/strata/artifact/gguf_reader.hpp`) reads the list, finds 0 on layer 3, and rejects the file.

**Suggested fix.** Treat a list that is all zeros as not given, and check it only when some entry is non-zero:

```diff
         int64_t block = 0;
-        for (int64_t l = 0; l < g.n_layers; ++l) {
+        // a list of zeros only is a converter's placeholder, not a layer split
+        bool given = false;
+        for (int64_t l = 0; l < g.n_layers; ++l) given = given || v->items[(size_t) l].u != 0;
+        for (int64_t l = 0; given && l < g.n_layers; ++l) {
             const int64_t r = (int64_t) v->items[(size_t) l].u;
```

With that change, 0.1.44 loads the IQ4_XS and IQ3_M files of that repository. Over two rounds it runs at the same speed as 0.1.43, and it passes the same checks: tool calls, fixed answers, a picture, and needles at 32K and 128K. A list with any non-zero entry is still checked as before.

**Setup:** Linux, RX 7900 XTX (gfx1100), ROCm 7.2.4, Strata v0.1.44 built from source. Nothing in 0.1.44.1 touches this check.

Happy to open a PR with this change if you prefer that.