Fix math RFT reference answers and clarify execution workflow

[vimalcha]

Summary

This carries the same seven-file changes as https://github.com/johnwu0604/fine-tuning/pull/17, including its latest documentation update, onto the current upstream main. Only the two recipe commits were transferred; unrelated repository changes are excluded.

• Correct 23 training and 6 validation reference answers in place, preserving prompts, record order, schema, and the 100/10 splits. Update canonical hashes and source provenance.
• Hardcode qwen3.6-35b-a3b as the notebook training model and remove the model environment setting from the blank template.
• Clarify model/artifact/deployment terminology, manual deployment handoff, job reuse/recovery, and monitoring timeout behavior using concise user instructions.

Validation

Passed locally on the upstream-based branch:

• 7 root documentation/model-eligibility tests.
• 96 selected managed workflow tests, including multimodal tests.
• 3 math validation tests.
• 10 local grader/adversarial cases and 110 boxed self-reference checks.
• Notebook schema/syntax, answer-only dataset edits, canonical hashes/byte counts, prompt isolation, blank template, output-free notebook, and diff checks.

Scope and limitations

• Exactly seven math-recipe files; no populated environment files, package-install artifacts, duplicate datasets, or runtime outputs.
• Grader implementation and training hyperparameters are unchanged. Text grading does not establish algebraic equivalence or comprehensive mathematical correctness; historical training metrics are not evidence for the current inputs.
• No Azure uploads, training, inference, deployment, or resource operations were performed for this PR. Cloud eligibility and held-out improvement are not claimed.