Port NInfer to RTX 4090 and Qwen3.6-27B

[jram4]

Summary

• port the inherited NInfer runtime from RTX 3090 sm_86 to RTX 4090 sm_89 on Linux
• specialize the compiled and registered product for Qwen3.6-27B
• retune Gated-DeltaNet dispatch for Ada's 128 SMs
• retain measured Q5 residual improvements and restore slower projection candidates to their inherited controls
• add bit-exact validation, curated benchmark evidence, public documentation, and maintenance workflows

Attribution and contribution boundary

This is a maintained port of Don-Chad/ninfer-3090, derived from Neroued/ninfer. MTP, CUDA Graphs, the .ninfer artifact format, C++ Engine API, CLI, and OpenAI/Anthropic-compatible server are inherited capabilities.

The contribution in this branch is the sm_89 architecture port, RTX 4090 GDN dispatch, retained Q5 optimization, 27B-only product specialization, tests, benchmark evidence, documentation, and public maintenance setup.

Verification

• Hardware qualification evidence: one RTX 4090, CUDA 13.0, GCC/G++ 13.4, Linux
• Stored result: 51.716 tok/s ordinary decode; 124.574 tok/s MTP-3
• Local portable suite: 20 passed (tests/artifact, tests/test_bench_matrix.py)
• Public-tree, secret, local-link, YAML, JSON/CSV, and whitespace checks: passed
• Provenance: seven commits based directly on upstream 464e2d6

Release note

This is a source-only public release. No model weights, binaries, builds, profiles, request logs, credentials, or raw Nsight traces are included. The stored GPU results are historical qualification evidence; no post-sanitization CUDA binary is claimed as validated.