[zhouaihui]

Outcome

Document two Draft experiments sharing Qwen3-8B Target features with separate learning rates, TensorBoard curves and checkpoints. Show how to resume from different completed steps when one experiment has already finished, and export each completed Draft. Include the walkthrough in source distributions.

Verification

• 62 shared-planning, runtime and recovery tests passed; lint, documentation links and wheel/sdist audits passed in an isolated Linux environment.
• Eight-GPU functional acceptance covered A 0→8 and B 0→4, followed by A 8→8 and B 4→8. Committed work, independent tracking and both native exports were checked.
• A child with no new training reports completed / incomplete_evidence, so the root reports incomplete_or_failed; the walkthrough documents the required state and cleanup checks. This does not establish quality or throughput.

Compatibility and provenance

Reuse the existing shared-target recipe and runtime. Exact resume clears both phase-transition flags in each complete experiment entry. No configuration schema, checkpoint or export format changes; no new upstream implementation is introduced.