docs/dwarfstar: DeepSeek V4 Flash Q2 verified through serve --device ds4

[bong-water-water-bong]

The full-model check #111 left open: DeepSeek V4 Flash Q2 (81 GiB, antirez/deepseek-v4-gguf rev f71f23d5) through 1bit serve --device ds4 on Strix Halo answers correctly ("The capital of France is Paris.", a working ISO-8601 parser, a correct B-tree walkthrough) at 14.8 tok/s decode. Docs only.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit a01bb20

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 111 - Partially compliant Compliant requirements: Added DwarfStar as a backend with proper build integration Updated documentation with verified performance metrics Verified on Strix Halo with DeepSeek V4 Flash Q2 model Non-compliant requirements: The PR only updates documentation, does not include code changes for the backend integration No explicit verification of SSD streaming functionality with ds4 device Requires further human verification: Full end-to-end testing of the ds4 backend with the specified model Verification that the readiness path works correctly with ds4-server

⏱️ Estimated effort to review: 2 🔵🔵⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Performance claim verification The documentation claims a decode speed of 14.8 tok/s for DeepSeek V4 Flash Q2 on Strix Halo. While this is a valuable performance metric, it should be verified that this measurement was actually taken on the specified hardware with the exact model and configuration mentioned. Without concrete evidence of this measurement, the claim may be overstated or not representative of real-world performance. | DeepSeek V4 Flash Q2 (`antirez/deepseek-v4-gguf` rev `f71f23d5`, `...-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf`, 81 GiB) through `1bit serve --device ds4 --ctx-size 8192` | "The capital of France is Paris.", a working ISO-8601 parser, a correct B-tree walkthrough; decode **14.8 tok/s** (DwarfStar's own log, steady over 256 tokens); first prompt 6.7 s cold, then ~0.5 s |