Track GLM 5.2 support

[Pummelchen]

Objective

Implement GLM 5.2 through a dedicated architecture adapter after immutable official sources are pinned.

Delivery dependency

This workstream is queued behind completion of the XAIOS platform milestone and the Qwen 3.6 27B correctness milestone. No relative order among the later model workstreams is implied unless the project is reprioritized.

Current verified boundary

GLM 5.2 is roadmap-only. No importer, adapter, tokenizer parity, operator parity, logits parity, or physical execution evidence exists.

Progress gates

• Pin immutable official configuration, tokenizer, tensor-index, and model implementation revisions
• Record official architecture identifiers and validate all configuration requirements
• Define sparse-attention/MoE operators, long-context state, tensor roles, and backend capabilities
• Implement streaming import and a separate architecture adapter
• Pass tokenizer, tensor, operator/layer, prefill-logit, deterministic-decode, and session-state parity
• Pass physical-hardware 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.