deepseek41 : add compressed attention runtime

[ajaxdude]

Summary

Stacked on the schema layer in halo-box#49 at 65fdc19831d30b12f06b981c3ca74dd18731f90e.

• load and strictly validate the published deepseek41.* metadata contract, including the 40-entry emitted compression layout and Engram header metadata
• add distinct DeepSeek V4.1 tensor ownership for KV sources [2,8,14,20] and index sources [2,8,14,20,24,28,32,36]
• add model-free V4.1 cache/state planning for ratio-0 raw attention, ratio-1 direct compression, ratio-2 two-token pooling, the 128-token raw ring, and source-owned compressed/index state
• add candidate block selection and propagation for layer 20 with block size 8, top 2048 blocks, causal filtering, stable ties, and forced retention of the final visible block
• add GGML graph primitives for ratio pooling, one shared raw-plus-compressed softmax, and final four-stream HC collapse through the BF16/RMSNorm/output path without output_hc_*
• add parameterized raw/compressed/index/carry/candidate/position/workspace memory accounting
• register a distinct DeepSeek V4.1 model and tensor contract using ATTN_KV_A_NORM, omitting V4-only output HC, compressor APE, indexer-compressor, TID2EID, and nextn families

Design: halo-box#48 and its refinement comment.

Runtime gate

The published model still requires disk-backed Engram decoding and routed-expert streaming. This layer validates tensor metadata and the four published Engram tensor names (engram_embd, engram_q_norm, engram_k_norm, engram_kv), then fails before payload mapping with an explicit dependency error. It does not claim that the published 365 GB GGUF can execute yet.

Validation

• test-deepseek41-runtime
• test-deepseek41-schema
• test-llama-archs
• git diff --check

No model payload was loaded or run. No performance claim is made.

Stack note

This PR targets the fork-only base branch jeromecoste-microsoft-deepseek-v41-support. It must be recreated or retargeted to halo-box/strix-llama.cpp after halo-box#49 merges.

Written by GPT-5.6 Sol. The implementation was validated only with local synthetic/model-free tests on macOS; no Strix Halo hardware or published-model inference was used.

[ajaxdude]

Stack repair update: merged the current schema base 6a473b7e942a2cc54f1e5bb2b2d5deacd247b888 normally. New attention head: 580986ee4a86ad2ec5a73d367c8f220e06377da8. Focused runtime and schema tests pass; test-llama-archs -a deepseek41 passes, and diff checks pass. The unfiltered local test-llama-archs run traps in unrelated Metal coverage after many non-DeepSeek architectures; the DeepSeek41 case remains intentionally skipped behind the dependency gate.