deepseek41 : activate scheduler-backed full graph

[ajaxdude]

Overview

Activates the complete scheduler-backed 40-layer DeepSeek V4.1 Flash graph on top of the accepted schema, attention, Engram, expert-store, and bounded expert-execution layers.

The model-specific memory context owns bounded raw 128-window state, source-owned ratio-2/ratio-1 compressed KV, index keys, compressor carry, layer-20 candidate propagation, sequence positions, and Engram history. State is prepared transactionally and committed only after successful graph execution, backend completion, runtime error consumption, and expert lease completion. Supported single-sequence semantics include reset, clear, full copy, keep, full removal, immediate last-ubatch rollback, and save/restore into empty destinations. Unsupported coupled/partial or older rollback operations fail explicitly before mutation.

The graph applies Engram before attention HC mixing at layers 1 and 14; implements ratio 0/1/2 raw and compressed attention with shared softmax, source maps, compressor publication, index/candidate selection, shared plus routed MoE execution, four-stream HC mixing, and the normal output norm/head; and preserves original expert IDs for routing/telemetry while using cache slot IDs only for expert GEMMs.

Depends on #6 and tracks halo-box#48.

Reconciliation

Final head: dd2fd630d03edf6e1bf42f82ed4f034fd89b5d26

Final tree: 99085493075106773be2e7724f07cf2c945a1fa8

dd2fd630d is a tree-identical provenance successor of merge commit 5e013725ea45486f95943a6ec12e1105171a1c41. The merge commit parents are frozen PR #7 head 4f1a4ebd74b58f1b22580b5c7f5ab976fdb99506 and exact accepted PR #6 head 83024f92db6b344b55f75103c43507788d853b0a. Excluded PR #6 candidates were not consumed.

Mandatory production fixes

• The final visible candidate block is physically excluded from ordinary top-(k-1) ranking and concatenated exactly once. k=1 returns only the final block. Executable all-+inf tests cover short k=1/k=2 inputs and 16,392 visible rows with 2,048 selected blocks, proving unique in-range IDs and final block 2,048 exactly once.
• Four-branch HC collapse remains BF16, then crosses an explicit F32 boundary before final RMSNorm. An executable CPU-supported graph continues through the output matrix and produces finite logits with expected value 32.0.

Model-free validation

• ctest --test-dir build-dsv41 --output-on-failure -I 27,40: 14/14 passed, including DeepSeek41 schema, Engram, expert/store, memory, runtime, no-allocation graph construction, generated architecture fixtures, recurrent rollback variants, and save/load state.
• ctest --test-dir build-dsv41 --output-on-failure -R '^test-state-restore-fragmented$': passed.
• PYTHONPATH=gguf-py python3 -m unittest gguf-py/tests/test_deepseek41_schema.py: 4/4 passed.
• CPU AddressSanitizer build with leak detection disabled because Apple ASan does not support it: 8/8 DeepSeek41/Engram/expert tests passed.
• test-llama-archs -a deepseek41 -s 1: passed. The metadata-only path rejects malformed present tensor dimensions, registers the external expert/Engram descriptors without loading their data, and reserves complete batch-64 and batch-1 graphs. Reported CPU compute workspace: 85.78 MiB; graph nodes: 7,413 / 6,858; graph splits: 636 / 1.
• test-llama-archs -a qwen4exp -s 1 and -a grovemoe -s 1: exited 0 with CPU, Accelerate, and Metal/device comparisons passing. Meta roundtrip remains the established test limitation (SKIP), not a DeepSeek regression.
• git diff --check, conflict-marker checks, ASCII-only changed-file validation, exact ancestry checks, and accepted PR deepseek41 : add bounded routed expert execution #6-only file preservation passed.

Measurements and unverified gaps

No performance measurement or real-model correctness claim is made. No model or GGUF data was loaded. No Strix host, Vulkan/RADV, ROCm, corpus/logit equivalence run, or benchmark was used. The calibrated GGUF is unchanged. A guarded metadata-only or no-allocation real-GGUF probe is structurally possible, but it was not run. Automatic host-memory fitting and wave-prefill remain out of scope.

Requirements

• I have read and agree with CONTRIBUTING.md.
• This change is Strix Halo specific; general upstream-bound work belongs in halo-box/llama.cpp.
• Agent-authored by GPT-5.6 Sol. The final head contains parseable Assisted-by, Copilot-Session, and Copilot Co-authored-by trailers.