#1491 bench: completed Flash Next tool-use results with charts and examples

本文

## Summary
Replaces closed #1456 with the completed twelve-variant evaluation and visual results.

All twelve requested model variants now have completed Tool-Eval-Bench results. This update adds Gyro-M and the three AP variants, completes the matrix, and includes two figures with short, paraphrased test examples.

Benchmark created by **[SeraphimSerapis](https://github.com/SeraphimSerapis)**: **[Tool-Eval-Bench](https://github.com/SeraphimSerapis/tool-eval-bench)**. Scenarios and scoring are unchanged.

![All twelve variants: short and standard scores](https://raw.githubusercontent.com/CC-David-CC/Strata-a5500/work/tool-eval-variants-20261007/bench/results/tool-eval-20261007/figures/variant-scores.png)

![Category scores and paraphrased test examples](https://raw.githubusercontent.com/CC-David-CC/Strata-a5500/work/tool-eval-variants-20261007/bench/results/tool-eval-20261007/figures/category-scores-and-examples.png)

## Added results
| Variant | Short /100 | Standard /100 | Scored cases |
|---|---:|---:|---:|
| Gyro-M | 93 | 86 | 68 |
| AP-Q4_K_XL | 97 | 85 | 69 |
| AP-IQ3_XXS | 90 | 89 | 69 |
| AP-IQ2_S | 93 | 88 | 69 |

## What changed
Completed report, traces, portable deployment settings, resource samples, conversion records and reproducible PNG/SVG figures under `bench/results/tool-eval-20261007`. No engine or benchmark scoring changes in this PR.

## Interpretation
- Six ISTA/Swift runs share the mainline build and RTX PRO 6000. AP uses the same GPU and settings, with the existing Q6 build option enabled for its embedding.
- Gyro uses patched agentionai Strata rc1; TC-45 is excluded because its endpoint does not enforce required tool use. These are not results from the newer Gyro merge branch.
- Q8 uses a separate three-P4 engine; its short score is extracted from the first 15 standard cases.
- Single trials; Swift/AP are modified variants. Structured-output API restrictions, stale mock timestamps and background transfers are documented. Examples are paraphrases, not model quotes; all tool actions are mocks.

This benchmark report is independent of the Responses/cache integration in #1489 and the Responses implementation in #1446.