ML training pipeline and flaky test fix

[KrisSimon]

Migrated from GitLab MR !225 (open)
feature/train-ml → main
Originally created: 2026-04-02
Author: Kris Simon
Labels: 0.10

Summary

• Add complete ML training pipeline for ARO-Coder (notebooks 00-21) covering corpus collection, knowledge extraction, fine-tuning, DPO, evaluation, and packaging
• Fix flaky RepositoryObserver test (20% failure rate) by increasing HTTP timeout for observer tests and adding a pump loop for async handler output
• Add streaming execution support (StreamExample)
• Various performance optimizations (regex cache, action name cache, EventBus O(1) removal, VerbSets caching, lexer string interning, lazy filter/map)
• Book updates for ARO 0.8.0 across all guides
• CI pipeline improvements for merge request events

Test plan

• Full test suite passes 20/20 runs (was 16/20 before flaky test fix)
• RepositoryObserver passes 30/30 isolated runs