Add nine-toothed-operator-dev skill for 2026 spring T3-1-1 by tusu-code

[tusu-code]

[2026 Spring][T3-1-1] tusu-code

GitHub PR title (CONTRIBUTING-compliant, imperative, capitalized, no trailing punctuation):
Add nine-toothed-operator-dev skill for 2026 spring T3-1-1 by tusu-code
Branch: spring-2026-tusu-code-t3-1-1 (competition-required spring-2026-<githubid>-t3-1-1, lowercase kebab-case)

Skill

• Name: nine-toothed-operator-dev
• Team: 韩博 / tusu-code (solo)
• Problem: T3-1-1 NineToothed Operator Development Skill

Scope

Applicable:

• Elementwise and broadcast operators
• Reduction and block operators
• Non-contiguous, stride, and offset-sensitive tasks
• Correctness tests against PyTorch or existing implementations
• Benchmark, generated source, AOT build, and failing-test diagnosis

Not applicable:

• NineToothed compiler core redesign
• Hidden evaluation answers or task-specific bypasses
• Online-only services or private credentials

Installation and Usage

Place the skill under:

skills/competition/nine-toothed-operator-dev/

Prompt:

Use the nine-toothed-operator-dev skill to implement this NineToothed operator task. Follow the required workflow, add correctness tests, run targeted validation, include benchmark evidence if performance-sensitive, and finish with changed files, commands, results, unsupported cases, and residual risk.

Self-Test Records

All GPU results from RTX 4090, torch 2.9.1+cu128, Triton 3.5.1, repo master c9ebd49.

Task	Category	Correctness	Benchmark	Notes
01-elementwise-add	Elementwise/broadcast	1 passed in 5.58s	1 passed in 9.80s; custom 0.0512 ms vs torch 0.0205 ms	Existing-op validation flow
02-softmax-reduction	Reduction/block	Core 1 passed in 1.86s	Custom 0.0543 ms vs torch 0.0215 ms	Examples benchmark blocked by Triton exact-match precheck; diagnosed as harness tolerance, not semantics
03-layout-stride-offset	Layout-sensitive	Root cause found (rank mismatch -> silent column-0 processing + empty_like contiguous fallback); both fixes allclose=True on GPU	N/A	Closed loop: repro -> root cause -> 2 fixes -> re-verify
04-performance-diagnosis	Performance/diagnosis	Core add/softmax passed	AOT re-verified with nvcc 12.8: tests/test_aot.py 11 passed, 1 skipped in 892.77s; full remaining suite 200 passed in 737.82s	Closed loop across two environments (runtime-only vs full toolkit image)
05-from-scratch-gelu	Elementwise (from scratch)	4/4 allclose vs torch GELU(tanh)	0.99x-1.05x vs PyTorch, 2^16-2^24	Includes real NameError failure-diagnosis loop (application globals not captured)
06-from-scratch-l2norm	Reduction/block (from scratch)	4/4 allclose, fp16/fp32, odd sizes	0.72x-0.80x of PyTorch time (20-28% faster, fused single kernel)	Real performance win with evidence
07-ab-comparison	Meta: with/without skill	A 8/8, B 9/9 (incl. non-contiguous view)	Both 0.50x-0.75x vs unfused torch baseline	Controlled A/B, checklist A 8.5/10 vs B 10/10, contamination caveat disclosed

Before/After Comparison

Measured A/B experiment (examples/07-ab-comparison/task.md): same agent, same repo snapshot, same GPU, same task text (softmax-temperature operator), condition A without the skill, condition B with the skill.

Dimension	Without Skill (A)	With Skill (B)	Improvement
Contract restatement	Partial (design notes only)	Full upfront contract section	Complete task-semantics discipline
Layout coverage	No non-contiguous test	Non-contiguous transposed-view test, passing	Covers the layout-sensitive grading axis
Output allocation	Not reasoned	torch.empty(shape) chosen over empty_like, citing documented pitfall	Pitfall-cited design decision
Scalar parameter	Tensor(0) (correct)	Tensor(0) with documented rejection of constexpr alternative	Reasoned trade-off
Risk documentation	Limitations list	Residual-risk section with one-line fallbacks	Reviewable failure plan
Checklist score	8.5/10	10/10	Measured gap is a lower bound (A had indirect exposure to skill-derived artifacts; disclosed)

Validation Commands

python scripts/check_submission.py --skill-dir .   # Structure check passed (structure/completeness only)
python -m pytest tests/test_scripts.py -q          # 9 passed (helper-script automation, stdlib only, no GPU)
python -m pytest tests/test_add.py -q              # 1 passed in 8.69s
python -m pytest tests/test_softmax.py -q          # 1 passed in 1.86s
python -m pytest tests/test_aot.py -q              # 11 passed, 1 skipped in 892.77s (nvcc 12.8)
python -m pytest tests/ -q --deselect tests/test_aot.py  # 200 passed in 737.82s

Main-Repo Style and Lint Checks

Run against the added skill scripts, per CONTRIBUTING.md (all clean on the latest commit):

$ ruff format --check scripts/
4 files already formatted

$ ruff check scripts/
All checks passed!

$ python scripts/check_contributing_style.py <skill scripts>
(no output; exit 0)

pytest output:

$ python -m pytest tests/test_aot.py -q     # RTX 4090, torch 2.9.1+cu128, Triton 3.5.1, nvcc 12.8
11 passed, 1 skipped in 892.77s (0:14:52)

$ python -m pytest tests/ -q --deselect tests/test_aot.py
200 passed in 737.82s (0:12:17)

Compliance

• HONOR_CODE.md signed
• REFERENCE.md completed
• No secrets or credentials
• No hidden answers
• No test bypasses
• External references disclosed

Attachments

• Competition reports (three documents, kept outside the skill runtime dir to keep the skill small and focused; each in .md + .pdf, packaged together with the skill and this PR description in the submission zip on the contest site):
  • Proposal: 九齿skill创新挑战_T3-1-1_提案.md (+ PDF) — dated 2026-06-26, planning only (motivation, design principles, workflow, self-test plan, milestones), no results
  • Mid-term report: 九齿skill创新挑战_T3-1-1_中期报告.md (+ PDF) — dated 2026-07-01, round-1 self-test checkpoint (elementwise/reduction passing; layout and AOT blockers found, root-cause pending)
  • Final report: 九齿skill创新挑战_T3-1-1_赛题报告.md (+ PDF) — dated 2026-07-02, all evidence loops closed (7 self-tests, correctness/benchmark, failure diagnosis, A/B comparison)
• Report template scaffolded by the skill: skills/competition/nine-toothed-operator-dev/reports/final_report_template.md
• Self-test records: skills/competition/nine-toothed-operator-dev/examples/*/task.md, tests/selftest_manifest.md
• Reproduction scripts and runbook: gpu-session/ (RUNBOOK.md, selftest_gelu.py, selftest_l2norm.py, verify_layout_fix.py, ab_test_*_impl.py)