Do Not Merge - Convert the runtime and transform stack to Qwen3.6 27B

[anupsv]

Summary

• replace the Gemma-specific model, harness, worker pinning, artifact defaults, and test suite with a Qwen-only text runtime for mlx-community/Qwen3.6-27B-4bit@c000ac2c… (qwen3_5_text internally)
• use the pinned library Qwen35TextModel as the production oracle with strict 64-layer hybrid-cache/config/1,847-tensor validation, explicit Qwen golden identity, explicit local baselines, and fail-closed worker/preflight contracts
• add a hard-disabled editable Qwen fast-path scaffold covering partial RoPE, gated GQA, Gated DeltaNet, SwiGLU blocks, transactional hybrid caches, and a last-token untied LM head, plus an opt-in real-checkpoint parity suite
• pin the complete Qwen checkpoint manifest and local setup/transform defaults while leaving correctness_prompts/**, the M5 benchmark workflow/configuration, and protected operator artifacts unchanged
• document the remaining operator handoff: provision Qwen on M5, generate Qwen goldens/GPQA data, calibrate thresholds and paired baseline identity, then activate the custom path only after real-checkpoint parity

Test plan

• swift build -c release
• swift test — 378 tests passed
• MLXFAST_RUN_MLX_RUNTIME_TESTS=1 swift test — 378 tests passed
• tiny deterministic library prefill/chunk/cache parity
• opt-in reference parity tests skip cleanly when MLXFAST_QWEN_REFERENCE_WEIGHTS_PATH is absent
• bash -n across shell scripts, JSON validation, and git diff --check
• no diff under correctness_prompts/**, .github/workflows/benchmark.yml, or ops/m5-bench/**
• real 15 GB transformed-checkpoint parity (operator follow-up; custom production gates remain false)

Important scope note

This prepares the codebase but intentionally does not make Qwen officially rankable yet. The protected M5 workflow and correctness artifacts remain Gemma-specific by request, so official Qwen runs must fail closed until the documented operator migration is completed.

Made with Cursor

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is disabled.}

[tenki-reviewer]

anupsv does not have permission to run Code Reviewer. You can manage permissions in your settings.