refactor: extract strategy classes from engine, fix_pipeline, and result_handler (fixes #518)

[mickume]

Summary

Extract collaborator classes from the three largest files in the codebase to improve readability and testability:

• BlockDecision / evaluate_review_blocking → engine/blocking.py (result_handler.py: 793 → 590 lines, −26%)
• SerialDispatcher / ParallelDispatcher → engine/dispatch.py (engine.py: 2051 → 1817 lines, −11%)
• CoderReviewerLoop → nightshift/coder_reviewer.py (fix_pipeline.py: 1237 → 980 lines, −21%)

Closes #518

Changes

File	Change
agent_fox/engine/blocking.py	New: blocking evaluation logic
agent_fox/engine/dispatch.py	New: serial/parallel dispatch strategies
agent_fox/nightshift/coder_reviewer.py	New: coder-reviewer retry loop
agent_fox/engine/result_handler.py	Removed blocking logic (−26%)
agent_fox/engine/engine.py	Dispatch delegates to new module (−11%)
agent_fox/nightshift/fix_pipeline.py	Loop delegates to new module (−21%)
4 test files	Updated import paths

Tests

• All 4292 existing tests pass unchanged
• 4 test files updated for import path changes only

Verification

• All existing tests pass: ✅
• Linter / formatter: ✅
• No regressions: ✅

---

Auto-generated by af-fix.