#1976 bench: RTX 5090 Laptop GPU follow-up for 0.1.43

Beschreibung

Follow-up to the RTX 5090 Laptop entries already on main (0.1.40.1, 0.1.40.3, 0.1.41): same laptop, model files, configuration and scripts, only the Strata version changed. Results only, no engine changes.

New folder: `bench/results/2026-10-11-community-rtx-5090-laptop-engine-0.1.43/`

**Median decode / prompt tok/s (Q2_0, 131K context, greedy, reasoning off, 256-token cap):**

| Prompt tokens | 0.1.41 | 0.1.43 |
| ---: | ---: | ---: |
| 4,096 | 142.5 / 1,949.8 | 140.5 / 1,906.9 |
| 32,768 | 141.5 / 2,888.3 | 141.5 / 2,871.2 |
| 128,000 | 127.8 / 2,776.8 | 131.6 / 2,767.5 |

- No measurable change on this machine. The route tail skip is on and the log counts skipped experts, but this card keeps 94–99% of decode expert lookups in VRAM, so there is little tail to skip. That fits the release notes, which expect the gain on 12–16 GB cards.
- Needles 6/6, coding check 10/10. Our private agent task suite shows no regression (122/132 vs 124 on 0.1.41, which is within spread).
- In case it helps others with their scripts: with reasoning on and a small `max_tokens`, #984 splits one request into two engine timing lines. Our own A/B script read only the last line until we noticed. Details are in the README.
- The source was downloaded by commit (`archive/<sha>.zip`), every file was checked against the git tree, and it was built locally for sm_120 with CUDA 13.3.

(#1462 and #1608 are still open, but their folders are already on main. Feel free to close them.)

🤖 Generated with [Claude Code](https://claude.com/claude-code)