[codex] Add benchmark suite, model onboarding pipeline, and recent CTR models

[shenweichen]

What changed

• add a reusable benchmark suite for single-task, multitask, and sequence models
• add the benchmarks.onboard workflow for discovery, scaffolding, wiring audits, verification, and documentation
• integrate OneTrans, FinalMLP, MaskNet, and WuKong, including serializable custom layers
• add a reusable C0-C6 correctness-testing contract:
  • deterministic fixtures
  • independent NumPy reference equations
  • finite and numerical gradients
  • semantic invariants such as masking and causality
  • prediction-equivalent weight and full-model serialization
• make scaffold --with-layer generate a correctness-test skeleton and make verify run it
• add model tests, benchmark tests, verification reports, docs, and contributor guidance

Why

Adding a model previously required manually updating several scattered registration, testing, benchmark, serialization, and documentation points. Existing smoke tests also proved that models could run, but not that paper equations, gradients, invariants, or serialization semantics were correct. This change makes onboarding repeatable and adds layered correctness evidence.

Impact

Contributors can scaffold and verify new CTR models through one CLI workflow. Users gain four recent models, reproducible benchmark tooling, and stronger regression protection across all existing models.

Validation

• dynamic onboarding audit: 33/33 models fully wired
• semantic correctness examples: 8 passed across reference equations, gradients, exact masks, causal invariance, and EDCN deterministic cases
• four new models plus multitask serialization contracts: 14 passed
• full suite first pass: 171 passed, with the stricter finite-output contract exposing one random-input EDCN NaN; deterministic fixture added and all four EDCN bridge configurations then passed
• scaffold templates compile and git diff --check passes

Large-sample effectiveness checks

Criteo_x1 2M, CPU, 1 epoch, batch 1024:

Model	AUC	LogLoss	Train
MaskNet	0.7908	0.4544	41.9s
DeepFM	0.7892	0.4559	32.6s
FinalMLP	0.7889	0.4562	32.6s
WuKong	0.7842	0.4614	61.5s

MovieLens-25M, 2.28M behavior samples, CPU, 1 epoch, batch 1024:

Model	AUC	LogLoss	Train
OneTrans	0.8056	0.5347	114.5s
DIN	0.8030	0.5375	12.4s

These benchmark numbers validate end-to-end effectiveness under a fixed protocol; equation-level correctness is covered separately by the semantic contracts above.

[codacy-production]

Not up to standards ⛔

🔴 Issues 12 high · 6 medium · 1 minor

Alerts:
⚠ 19 issues (≤ 0 issues of at least minor severity)

Results:
19 new issues

Category	Results
Security	6 medium 1 minor 12 high

View in Codacy

🟢 Metrics 586 complexity · 14 duplication

Metric	Results
Complexity	586
Duplication	14

View in Codacy

_{NEW Get contextual insights on your PRs based on Codacy's metrics, along with PR and Jira context, without leaving GitHub. Enable AI reviewer
TIP This summary will be updated as you push new changes.}