Refactor /referee2 into mode-specific audit protocols

[nsmiller2501]

Summary

This PR is a substantial redesign of /referee2. It turns the skill into a short mode router and moves the actual audit protocols into mode-specific reference files.

What changed

• Splits shared Referee 2 persona/report conventions into referee2.md.
• Adds deck.md for deck audits and code.md for empirical code audits.
• Keeps SKILL.md as a small mode router for deck vs code.
• Adds a tainted-session catch: if the parent session has already touched the target project, it must either cancel or act only as an orchestrator for fresh role-specific subagents.
• Reworks code audit around a specification bottleneck: Agent 0 checks comment/code and scope readiness; Agent A writes a plain-English/math spec from executable behavior; Agents B/C replicate from that spec and sealed expected outputs.
• Adds materiality classifications, override ledger, resume behavior for partial audits, and sequential-by-default fanout to reduce usage-cap risk.
• Files reports and audit artifacts under correspondence/referee2/ and code/replication/.
• Updates the README to explain the new role-subagent protocol and relationship to /blindspot.

Why

The main motivation is to preserve the independence of the audit. Reading original code and then directly translating it into another language can turn replication into syntax translation. The spec bottleneck separates intent extraction from independent implementation, making cross-language disagreements more informative.

Testing

• Reviewed the branch diff against the PR notes.
• Beta-test notes for this redesign are in the project planning artifacts (skills updates PR/), especially the referee2 code-audit beta spec.
• No automated tests were run; this is a skill/protocol redesign.