security: audit input handling and document security model

[emesal]

M6 work item. Harden chibi against malicious input; verify tein sandbox boundaries post-M2.5.

Audit areas:

• plugin output sanitization
• path traversal protection (verify unified path resolver covers all cases post-fix: unified path resolver for all file-accessing builtin tools #192)
• input validation for all CLI flags
• config file parsing safety (toml injection, unexpected field types)
• tool argument validation before dispatch
• tein sandbox boundary integrity (ensure sandboxed contexts cannot escape)

Deliverables:

• security audit notes committed to docs/security.md
• fuzz tests for at least the config parser and tool argument deserialisation
• any discovered vulnerabilities fixed
• safety limits added where missing (max transcript size, max plugin output size, etc.)

Depends on: M2.5 (tein synthesis, ToolRegistry) — sandbox boundaries only meaningful once tein tools exist