docs: align pipeline and dataset guidance

[bingran-you]

Summary

• Audit every project-facing document and update the 18 files that contain pipeline, dataset, compatibility, security, or evidence claims.
• Make the executable contract consistent everywhere: Qwen3.5-9B student, Qwen3.5-397B-A17B teacher, OpenCode for teacher/evaluation/GRPO rollouts, one-epoch bf16 LoRA SFT, and one-epoch TRL LoRA GRPO.
• Pin the public examples to the current 2,238-task train and 366-task eval repositories, while separating those public references from participant-corpus training and the sealed competition eval set.
• Reclassify older Qwen3-4B, OpenEnv, soccer, and HF Jobs runs as historical evidence and scope the 16x14 result as exploratory same-domain uplift rather than competition generalization.
• Document the canary provenance boundary, OpenEnv exposure requirements, HF Jobs topology gap, and private task-ID redaction rules.

This is the smallest appropriate change because executable code and configs already implement the target pipeline; the remaining drift was in repository documentation.

Validation

• Existing task checks pass when relevant
• Pipeline tests and documented CLI checks pass when relevant
• No credentials, private eval content, checkpoints, or raw provider dumps are included
• Dataset, model, and external dependency revisions are pinned when relevant
• Public docs and evidence boundaries are updated when behavior or claims change
• Compatibility changes update docs/architecture-status.md and retain existing contract coverage

Detailed checks:

• 238 pipeline tests passed
• Ruff and Python compilation passed for the pipeline package
• Starting-kit structural checks and submission checks passed
• Seven checked-in recipes validated; the full plan resolved to 2,238 train tasks, 366 eval tasks, Qwen3.5-9B, the Qwen3.5-397B-A17B teacher, and one epoch each for SFT and GRPO
• Qwen3.5 HF Job bundle dry-run passed
• Markdownlint passed on all 18 changed docs
• Modified-doc external and relative link checks passed
• Live Hugging Face verification matched all four pinned model/dataset SHAs; datasets contained 2,238/366 task.md packages and zero task.toml files
• Diff whitespace and credential-pattern scans passed

Spend and external systems

No spendful validation. This PR uses the already-preserved public canary evidence and no GPU, provider, Daytona, or paid HF Job was launched. A paid HF Jobs run for the current Docker/OpenCode/two-GPU topology remains an explicit manual validation item.