fix(amd): repair dense Qwen GGUF swapping and anonymize deployment reports

[dbourdea]

Repairs mixed-format dense Qwen attention and GDN loading, preserves packed output blocks with activation reordering, and fixes the dense tokenizer mapping. Anonymizes AMD documentation, JSON evidence, publication author/contact metadata, and the regenerated six-page review PDF. Benchmark host safety checks now require an explicit operator-supplied hostname. Validation: 21 model unit tests, 32 benchmark/privacy/reproducibility tests, exact Qwen3.6 and Qwen3.8 CPU/meta contracts and tokenizer checks, and two real Qwen3.6-to-Qwen3.8-to-Qwen3.6 swap runs through a pinned unmodified llama-swap binary. The extended run additionally passed same-model and conflicting-model concurrency, streamed usage validation, and TTL idle eviction. The protected service was restored and its completion verified. Live integration also requires the /ready endpoint from PR #1. Detailed evidence, checksums, and limits are in docs/qwen-swap-validation.md. Public test-host name: GMKtek EVO-X2. Raw operational evidence remains private. Historical commits and previously distributed artifacts are not rewritten. Bounded live cancellation and native loader-failure recovery now passed, including protected-service restoration and generation checks. This remains a draft for maintainer review; broader quality and endurance are not claimed. FreeToken is the only PR target; no llama.cpp or llama-swap submission is included.

[dbourdea]

The newly approved live window passed real Qwen3.6 cancellation through unmodified llama-swap, followed by Qwen3.6-to-Qwen3.8-to-Qwen3.6 streaming, concurrent routing, and idle unloading. Native daemon recovery passed a real invalid-GGUF loader failure, restored Qwen3.6 readiness, and verified streamed generation. Both phases restored and verified the protected llama.cpp workload. Final checks found no test listeners or FreeToken multiprocessing workers. Updated docs/qwen-swap-validation.md distinguishes these bounded results from broad quality and endurance claims. Integration and recovery harnesses remain in PR #1.