feat: add train-model and demonstrate-value skills, exponential backoff in babysit-training

[lawwu]

Closes #2, #3, #5

Summary

• Add exponential back off to babysit-training #2 — Exponential backoff in babysit-training: adds --max-interval and --no-backoff flags; interval grows 1.5× per uneventful poll (capped at --max-interval, default 300s) and resets on anomaly, >5% loss delta, new epoch, or Vertex AI state change; progress output now shows next poll in Xs
• Add an explicit training skill and also add it to orchestrate_e2e and benchmark_e2e #3 — train-model skill: new action skill that parses and launches a training command, applies HP overrides, delegates monitoring to babysit-training, enforces early stopping, tracks checkpoints, and invokes check-failed-run on failure; orchestrate-e2e gate 6 now calls train-model instead of running training inline
• add skill /demonstrate-value #5 — demonstrate-value skill: new post-eval skill that reads check-eval.json, explain-model.json, and build-baseline.json, translates metrics into business-language impact estimates, and renders a stakeholder-ready HTML presentation via showboat; added as an optional step in orchestrate-e2e after gate 8

Files changed

• skills/babysit-training/SKILL.md — backoff logic, new flags, updated example
• skills/train-model/SKILL.md — new skill
• skills/demonstrate-value/SKILL.md — new skill
• skills/orchestrate-e2e/SKILL.md — gate 6 → train-model; optional demonstrate-value step
• skills/orchestrate-e2e/references/lifecycle-gates.md — gate 6 evidence updated
• skills/orchestrate-e2e/references/artifact-contract.md — gate numbering fixed; train-model.json and demonstrate-value artifacts added
• skills/benchmark-e2e/references/skill-matrix.md + modes.md — stage 6 updated to train-model
• references/schemas.md — train-model and demonstrate-value schemas; orchestrate-e2e gate_name enum updated
• references/vocabulary.md — Terminal State enum added
• README.md — train-model and demonstrate-value rows added (14 skills total)
• plugin.json — bumped to 0.2.0

Test plan

• claude --plugin-dir ./plugins/agentic-ml — verify train-model and demonstrate-value appear in skill list
• Invoke /train-model with a toy training command and confirm it delegates to babysit-training
• Invoke /babysit-training with --interval 5 --max-interval 30 on a long-running process and confirm interval grows and resets on anomaly
• Invoke /demonstrate-value with existing check-eval.json and confirm HTML is generated