Track Qwen 3.6 27B support

[Pummelchen]

Objective

Complete Qwen 3.6 27B as the primary real-model correctness target before model-specific performance claims.

Delivery dependency

This is delivery workstream 2. Implementation starts only after the XAIOS platform milestone and tracker #14 meet their completion criteria. Until then, the Project status is Blocked, not In Progress.

Current verified boundary

Qwen 3.6 27B is interface-only. The model-v2 package and architecture/backend interfaces exist, but official tokenizer parity, transformer execution, logits parity, deterministic decode parity, and physical-hardware validation are incomplete.

Progress gates

• Stable model-v2 package skeleton and scalar architecture/backend canary
• Pin immutable official configuration, tokenizer, SafeTensors index, and model implementation revisions
• Implement bounded-memory streaming import and package-owned tokenizer parity corpus
• Implement the exact configured hybrid layer plan, including attention/recurrent/convolution state where required
• Pass embedding, complete-layer, prefill-logit, and at least 32 deterministic decode-step golden gates
• Pass session save/reload continuity and unsupported-field rejection
• Add no-expand scalar/NEON/AVX2 and capability-gated optimized backends with differential tests
• Pass physical Apple and Xeon correctness and benchmark-evidence gates

Status source

Execution status is tracked in the XAIOS GitHub Project. Support claims remain governed by docs/MODEL-SUPPORT.json.