Fix remaining agent workflow and training validation gaps

[Harshith Arun Kumar (harshith-arunkumar)]

Summary

Follow-up from current main after #46; this does not repeat the merged paid-run approval cleanup.

• Forward Tulu3's existing fail_on_truncation and model_context_length options to the shared renderer validation for both training and evaluation. Default truncation behavior stays unchanged.
• Stop managed retries when backoff reaches or exceeds the retry deadline, preserving the original exception. In-flight operations are not interrupted.
• Remove the remaining routine-setup approval ambiguity and make the quality troubleshooting checklist conditional on stalled/regressing results.
• Correct clone/support links to microsoft-foundry/fine-tuning, contribution guidance to main, and broken sample-dataset navigation. Distinguish the small NL-to-Python sample from the maintained demo's larger preserved dataset and clarify historical source notebooks are not included.
• Add regression coverage and run the navigation checks for changes to the affected documentation subtrees.

Validation

New regressions reproduced the defects before the fixes: five Tulu3 cases, two retry-deadline cases, six stale upstream references, and two missing local dataset links.

Current-branch offline validation on Windows/Python 3.13:

• Root documentation: 9 passed.
• Full interactive suite with external networking denied and published SDK 1.0.0b1: 1,952 passed, 11 skipped (uncached tokenizer/platform-specific checks).
• Managed shared, multimodal, and agentic contract suites: 125 passed.
• All 237 Python files, 16 notebooks, 27 JSON files, 57 JSONL files, and 113 Markdown files passed syntax/schema/output/navigation checks as applicable.
• Interactive dependency check, editor diagnostics, and whitespace checks passed.

No live Azure operations or whole-notebook execution. Canonical datasets, hashes, notebook code, dependency pins, preview access requirements, sandbox safeguards, and owned-resource/deletion boundaries remain unchanged. Local tests used existing environments; CI covers the supported OS/Python matrix.