Add Qwen3.8-27B + Muse Glimmer 30B llama.cpp configs (single 3090)

[Isaac-opz]

Summary

Adds two new model directories with validated llama.cpp configs for single RTX 3090, plus a reference doc for native Windows inference. No changes to existing files.

• Qwen3.8-27B (models/qwen3.8-27b/): dense 27B, hybrid GDN+FA, native 262K ctx, MTP head embedded, vision + tools. Four profiles validated on llama.cpp b10435 (native Windows): mtp-170k (~40–44 t/s), max-262k (~30–33 t/s), think-262k, balanced-128k.
• Muse Glimmer 30B (models/muse-glimmer-30b/): Meta Superintelligence Lab dense 30B (SWA+GQA), DFlash n=15 at 131K: 47 t/s post-sweep, native tool calling + vision. Validated on WSL2 (build 10349).
• docs/WINDOWS_NATIVE.md: reference notes for llama-server.exe on native Windows (--mlock, ubatch sizing, gpu-layers, parallel, cold start, process management patterns).

Single-rig data throughout. Each model has a CHANGELOG with measured numbers and explicit N/A justifications for the verify/soak/bench gates that don't apply to the non-Docker path. Happy to re-run under your harness or restructure to match the catalog registry format if preferred.

Type of change

• New compose variant (models/<model>/<engine>/compose/docker-compose.<name>.yml)
• New patch / sidecar (models/<model>/<engine>/patches/)
• New script or tool (scripts/, tools/)
• New model (models/<new-model>/)
• Doc-only / typo
• Other (describe) — reference YAMLs (not Docker Compose) + native Windows doc

Verification

• Full rig + validation report attached
• Profile header complete (new/changed compose)
• BENCHMARKS row added
• CHANGELOG entry added in models/<model>/CHANGELOG.md.

N/A justifications (if any boxes above are unchecked)

• report.sh / verify-full / verify-stress / soak-continuous: N/A — these target Docker containers. The configs here are native Windows (llama-server.exe) and WSL2 host builds with no container to smoke-test. Validation is sustained production use: 7+ days continuous agentic workloads (Hermes Agent multi-turn tool calling, long-context RAG, vision) per profile without degradation or VRAM creep.
• Profile header: N/A — these are reference YAMLs documenting CLI invocations, not Docker Compose files. No test-compose-status-drift gate applies. Happy to restructure into proper compose + registry entries if you want them in the catalog.
• BENCHMARKS row: deferred — single-rig data (one RTX 3090, one OS). Numbers are in each model's README + CHANGELOG with full hardware context. Happy to add a BENCHMARKS row once you confirm the column shape you'd like, or if another rig confirms the numbers.

Cross-links

• Closes #
• Related upstream: llama.cpp MTP spec-decode (merged), DFlash draft KV quantization issue #25725 (fix #25823 in build 10349)