feat: multi-pass review with parallel specialized personas

Summary

Run multiple review passes in parallel — each with a focused persona (e.g., security, correctness, style) — then merge and deduplicate the results into a single consolidated comment.

A single prompt asking the LLM to review for "everything" produces diluted results:

• Attention splitting: when asked to check security, performance, style, and correctness simultaneously, the model tends to surface only the most obvious issues in each category rather than going deep on any
• Prompt overload: a system prompt covering all concerns becomes long and contradictory ("be concise" + "be thorough about edge cases")
• Inconsistent coverage: some PRs get security feedback, others don't — it depends on what the model "notices" in a single pass

Human review teams solve this with role specialization (security reviewer, domain expert, style nit-picker). We should mirror that.

Proposed Implementation

1. Persona definitions: Each persona is a prompt file in the prompts/ directory (already partially supported via --persona). Define default set: correctness, security, style.
2. Parallel execution: Run N persona passes concurrently against the same diff context. Each pass uses the same files/budget but a different system prompt tuned for its focus area.
3. Result merging: Collect all comments, deduplicate by (file, line, semantic similarity), and assign severity labels from each persona's perspective.
4. Consolidated output: Post a single review comment organized by file, with severity indicators and persona attribution (e.g., "🔒 security: ...").

Start by reading the existing persona support in the prompts/ directory and the --persona entry point. Trace how the diff context is passed to a review and how the consolidated comment is posted. Done means correctness, security, and style passes run concurrently, their comments are deduplicated with severity and persona attribution, and one organized review comment is posted.