[jayleaton]

What this changes

DeepSeek-V4.1 serving needs prepared assets: the token map, RoPE tables, signatures, the AOT needs list and SM121 cubins. Producing them meant booting a CUDA model. This installs the token-map exporter and the real AOT planner as a CPU tool, dsv41-assets, that needs no GPU. It parses the complete release metadata for preparation, while serving keeps its checkpoint representation refusals.

A second commit lets the same tool list Engram shard paths from an index-only source. It keeps repository-relative paths and requires an explicit prefix for bare filenames. Missing or unsafe index data is refused with an actionable message. No model shards or device are needed.

Docs: docs/recipes/deepseek-v41-cuda.md describes the CPU preparation steps.

Issue: none; this is tooling for the existing experimental family. Stacked on: nothing (based on main 3f0a715c4; the 28 later main commits, to f350cd3ea, touch none of its files).

Receipt

• Environment, with the TensorFold commit, MLX or PyTorch version, machine and GPU, checkpoint and revision: head ddde40de8812 on main 3f0a715c4; Zig 0.17.0; host checks on a Linux aarch64 DGX Spark (GB10, CUDA 13.0). Real-pack boot on two GB10s, TP=2 over RoCE, checkpoint jayleaton/DeepSeek-V4.1-Flash-1017-EXL3-2.8bpw revision 859396591c34, TF_EXPERIMENTAL=deepseek.
• Exactness: n/a; this is a preparation tool, and serving arithmetic does not change.
• Decode speed before and after: n/a.
• Prompt speed before and after: n/a.
• Evidence: from the public 1017 config and metadata, the tool produces the canonical token map, RoPE, signatures, needs and an SM121 cubin. Host checks cover CLI admission, a refusal to select a GPU operation, the public 1017 config fixture, and Engram index refusals.
• Tests run, with pass, fail and skip counts:

  zig fmt --check build.zig zig                      exit 0
  git diff --check 3f0a715c4..HEAD                    exit 0
  python3 tools/zig/lean_check.py                     base 425 problems, head 425, new 0 (lean_new.py)
  zig build test-cuda-host --summary all              exit 0, 271/271 steps, 3906/3915 tests passed, 9 skipped (with test-server-cpu)
  zig build test-dsv41-shared test-dsv41-state test-dsv41-forward test-dsv41-kv test-dsv41-device-host test-dsv41-draft test-dsv41-cuda-host test-draft-dsv41-host test-dsv41-assets-host test-dsv41-token-map dsv41-assets  exit 0, 63/63 steps, 552/554 tests passed, 2 skipped
  

• Not run, and why: Metal; macOS host build of the tool (no Mac available).
• Overlap: atc overlap reports a one-line conflict in build.zig with draft lfm3: serve LFM3-1.7B on CUDA: NVFP4 and FP8 weight-only, MLX 4-bit (stacked on #1513, #1492 and #1503) #1508 and with lfm3 CUDA: serve the block FP8 export weight-only and the MLX 4-bit export as stored (stacked on #1508) #1634, which is stacked on it. Each PR adds its own family host-test line, and the merge takes both.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off.
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (Not measured for this change alone.)
• No precision traded for speed. If the bits change, the description says which sums change and why.
• A new or changed GPU kernel is registered with the kernel contract (docs/kernel-contract.md); each backend's SUMMARY line is in the receipt. (No kernel changes.)
• Prompt processing is no slower than the last release. (Tooling only.)
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (No reverted-fix control: the tool and its tests are new.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.