feat(agent): add in-pipeline pre-PR self-review phase

[nizar-lahlali]

Summary

Add an optional self-review phase to the agent pipeline that runs between agent execution and post-hooks. The LLM critiques its own cumulative diff before the PR is created, improving first-pass PR quality by catching bugs, style issues, edge cases, and test gaps that would otherwise go straight to human review.

Motivation

Post-hooks in pipeline.py run deterministic checks (build/lint) after the agent finishes, but the LLM never sees its own cumulative diff before the PR is created. Adding a self-review phase where the model critiques its own diff and iterates on fixes improves first-pass PR quality.

Acceptance Criteria

• New agent/src/self_review.py module with run_self_review() orchestration function
• New agent/src/prompts/self_review.py with focused review prompt template
• TaskConfig extended with self_review_enabled (default False) and self_review_max_turns (default 5) fields
• Fields threaded through build_config(), get_config(), server.py, and pipeline.py
• Self-review phase inserted between cancel-check and post-hooks in pipeline
• Skip conditions: disabled, pr_review task type, no diff, no remaining turns/budget
• Fail-open: self-review errors do not block PR creation
• Uses remaining turns from original max_turns allocation (capped at self_review_max_turns)
• Progress milestones: self_review_started, self_review_complete
• Unit tests covering all skip conditions, diff helpers, and happy path
• Feature is opt-in (off by default) — existing tests pass without modification

Design

See the approved plan in the implementation PR for full design details including:

• Second run_agent() call with fresh review context
• Diff truncation at hunk boundary (60k char cap)
• Budget/turn computation from remaining allocation
• Fail-open error handling

Task Type

new_task