[Feature Request] Multi-pass review with parallel persona analysis

Problem

The current review pipeline makes a single LLM call with all context pre-assembled. This one-shot approach has several limitations:

• No self-correction — the LLM can't verify its own findings
• Hallucinated line numbers go undetected
• No understanding of PR intent before reviewing code
• A single generalist prompt spreads attention thin across security, correctness, and performance concerns
• False positive rate is higher than necessary

Proposal

Replace the single-call pipeline with a three-pass architecture that combines sequential refinement with parallel specialized analysis.

Pass 1: Summarize (single call, cheap)

Lightweight call with PR description + diffstat (file names and status only, no patches). The LLM produces a 2-3 sentence summary of what the PR does and why. This summary
feeds into Pass 2 as intent context.

• Input: PR title, description, file list with status
• Output: 2-3 sentence summary of intent
• Token cost: ~1k input, ~200 output

Pass 2: Parallel persona reviews

Run 2-3 specialized reviewers concurrently, each with the summary from Pass 1 plus the full file patches:

Each persona gets the full token budget for its domain — no competition between concerns. Personas run as goroutines for wall-clock parallelism.

The existing persona infrastructure (internal/config/assets/prompts/) already supports this — critical-strike.md is effectively the security persona.

Pass 3: Self-check and merge

All findings from Pass 2 are collected and sent to a final validation call:

1. Deduplicate — remove findings flagged by multiple personas
2. Verify — does each line number match the actual diff? Is the issue real given PR intent from Pass 1?
3. Rate confidence — high / medium / low per finding
4. Filter — post only high-confidence findings; medium with caveat; drop low

• Temperature: 0.1 (deterministic, conservative)
• This pass alone should eliminate 30-40% of false positives