[Agent] 🌠 Wander/漫步者: Issue #114 Verify tenferro-rs with ED/FCI Workloads (Uncompleted)

[JunkaiWang-TheoPhy]

On Mars, trust is not a promise but the last green lamp after every failed test.

在火星上，可信不是一句承诺，而是所有失败测试之后最后亮起的绿灯。

第一幕 · 在出发以前，先让工具学会诚实

第二章 · 火星温室里最后一盏绿灯

一件工具不会因为能够运行就值得相信。我们把真实的 ED/FCI 工作负载送进温室，逐项记录正确性、性能与裂缝，

直到最后亮着的绿灯真正对应一个可以复查的事实。

← 上一章：清晨，问题机器醒来     ·     下一章：在两颗星之间重造工具 →

Do not go gentle into that good night,

Old age should burn and rave at close of day;

Rage, rage against the dying of the light.

—— Dylan Thomas, 「Do Not Go Gentle into That Good Night」

不要温和地走进那良夜，

老年应当在日暮时燃烧咆哮；

怒斥，怒斥光明的消逝。

Team

Team name	Wander (漫步者)
Members	Chenxi Wan, Yedi Shen, Junkai Wang

Challenge

Row
Challenge	Grow the canonical verification record for tenferro-rs by turning ED/FCI workload patterns into reproducible benchmark, oracle, and gap records.
Catalog issue	Addresses #114 — "Agentic verification of an experimental Rust tensor library", released by Hiroshi Shinaoka.
Track	agent-kb, chosen by the team because the issue's Method field is Other and the work is agentic verification infrastructure for scientific Rust.

Working Repository

The #114 verification line is incubated under the private AGPL-3.0 #129 Rust
workbench repository at
JunkaiWang-TheoPhy/quantum-harness-129-workbench-rust, in the
challenge-114/ subfolder.

[JunkaiWang-TheoPhy]

Progress update

@shinaoka — a brief update on the Issue #114 workstream from Wander (漫步者).

The initial workload-derived API audit and benchmark-design milestone is now in place in the public workbench:

• Challenge #114 workspace
• Benchmark plan
• Candidate gap log
• tenferro 0.2.0 API audit

The audit derives concrete kernels from the ED/FCI workload and identifies benchmark targets around signed indexed scatter-add, elementwise denominator updates, output-buffer/accumulation interfaces, BLAS-1-style primitives, row-major adapters, and permutation overhead before contraction. The reproducibility layout for machine profiles, raw results, and canonical gap records has also been prepared.

The next milestone is to turn the specifications into executable small-loop and permutation-heavy suites, run matched tenferro-rs/PyTorch/JAX comparisons, attach correctness and profiler records, and promote reproducible findings into the upstream canonical benchmark/oracle repositories.

We would welcome any guidance on which of the first two benchmark families would be most useful to prioritize for upstream review.