release: r17 TR3-3.36 i8_hier + exact paired-FC2 image

[yatesdr]

What

Publishes the reproducible packaging for the qualified GLM-5.2 TR3-3.36 production candidate:

• immutable r17-derived GHCR image and digest-pinned TP4/DCP4 Compose;
• topology-specific four-rank i8_hier transport;
• bounded block-64 mixed-K3/K4 prefill with deterministic paired-M8 FC2;
• exact installed-file and source-provenance manifests.

Raw CN4 host logs are intentionally excluded from this public packaging branch.

Why

The faster block-64 FC2 reduction changed prefill logits and missed the KLD ceiling. Captured-input proofs isolated the difference to FC2. The paired-M8 implementation restores the original deterministic reduction order while retaining bounded block-64 route planning and weight reuse.

Qualified result

• prescribed KLD: 0.07776700789368204 (exact reference parity)
• cold 64k prefill: 1,737.11 tok/s, cached_tokens=0 (operator-accepted gate)
• MTP0/C1: 38.96 tok/s
• MTP3/C16: 234.05 aggregate tok/s
• MTP0/480k GPU KV pool: 609,024 tokens
• cold 350k needle: exact 738216

Published image

ghcr.io/yatesdr/glm52-serve@sha256:f4543b83de1d342463ebea1863ea0f4c23167e69b4e0e93938c3f6c07634bcbf

Code review

• i8_hier: [PCIe] add group-consistent i8_hier DCP transport b12x#1
• paired-M8 FC2: [EXL3] preserve mixed-prefill numerics with paired M8 FC2 b12x#2
• vLLM planner: [EXL3] keep bounded mixed-K prefill on one grid vllm-opt#1