v0.6.0: Security Hardening

Phase 1 Security Hardening (Items 1-7 — COMPLETE)

#	Item	Status
1	API Key Exhaustion — AllKeysExhausted with cloud→local fallback	✅
2	Auth/Authorization — ToolAuthError, dangerous flag, --auth-token CLI	✅
3	Input Sanitization — control tokens, path traversal, shell injection, MoA	✅
4	Rate-limit Hardening — per-key token buckets, jittered backoff, GLM-5.2 default	✅
5	Sandbox Tightening — command allowlist/blocklist, env scrubbing enforced	✅
6	Secret Redaction — secrets stripped from child env, stdout/stderr, and audit log	✅
7	Subagent Isolation — child env sanitized, parent env restored after run	✅

Stats

• 463 tests passing (was 432, +31 new security tests)
• 16 source modules, 20 test files
• Default model: z-ai/glm-5.2 (NVIDIA NIM)
• Local fallback: Qwen2.5-Coder-7B-Instruct-IQ3_XS (3.1GB, tool-calling capable)

Local Model Upgrade

• Switched from Gemma-4-E2B → Qwen2.5-Coder-7B-Instruct-IQ3_XS (Bartowski)
• Tool calling works natively via llama-server grammar
• 65K context, Q8_0 KV cache, 28 GPU layers (8GB M1 safe)
• Updated launchd plist + Hermes-Lite config defaults