feat: Squad Arena — AI adversarial review mode

[tt-a1i]

Summary

Add an Arena mode where two AI agents debate, attack, and defend code or technical proposals, with a judge agent scoring the results.

Usage

# Terminal 1: Attacker — finds bugs, vulnerabilities, and flaws
/squad red-team

# Terminal 2: Defender — fixes issues and counters arguments  
/squad blue-team

# Terminal 3: Judge — scores rounds and declares winner
/squad judge

Use Cases

• Code security audit: Red team finds vulnerabilities, blue team patches them
• Architecture debate: Two agents argue pros/cons of different approaches
• Algorithm competition: Who writes faster/cleaner code for the same problem

Why

Adversarial review catches significantly more issues than single-agent review. This mode makes multi-agent collaboration more engaging and produces higher-quality output through structured debate.

Implementation

Phase 1 (minimal): Add 3 new role templates (red-team.md, blue-team.md, judge.md) under .squad/roles/. Zero code changes — just prompt engineering.

Phase 2 (optional): Add squad arena subcommand that:

• Auto-initializes a 3-agent arena session
• Structures rounds (attack → defend → judge)
• Tracks scores and generates a summary report

Labels

enhancement, good first issue

[tt-a1i]

Phase 1 can stay very small here.

The current role loading already supports this pattern without new runtime code: after squad init, users can add project-local role prompts under .squad/roles/, for example:

.squad/roles/red-team.md
.squad/roles/blue-team.md
.squad/roles/judge.md

Then these work as normal roles:

/squad red-team
/squad blue-team
/squad judge

So the minimal built-in version would just make those three prompts official defaults:

• add red-team, blue-team, and judge to the built-in role list
• add bundled prompt templates for each role
• let squad init --refresh-roles write/update them like the existing manager/worker/inspector roles

No squad arena subcommand is needed for the first step. That can remain a later Phase 2 for automatic round orchestration, scoring, and reports.