[Fix] Fix runtime bugs

[fanqiNO1]

🚀 Desktop Semantic Tools Go Live

• Tool-Loop Injection: 17 cua-driver semantic desktop tools (observe_ui, click, type_text, open_url, list_apps, …) merge into the unified loop's per-turn catalog whenever the bridge carries them — auto-disappearing when perception is offline.
• Progressive Disclosure: Desktop is a non-core capability_expand category; CORE turns pay one index line, expansion on demand.
• Approval Gating: Mutating desktop tools go through the existing ApprovalOrchestrator (fail-closed); observation tools run directly.

🛠️ Runtime Consistency

• Cache-Key Fix: Unified catalog cache keyed on (bridge identity, registry size), so tools registered after engine construction (session_search) no longer vanish from prompts.
• Live /tools: TUI tool view renders the live catalog; desktop tools show under a desktop group when online.

🧪 Test Stability

• Deterministic Journeys: LEAPFLOW_DESKTOP_TOOLS_ENABLED=0 in the journey harness keeps cassette fingerprints stable — all six journeys green, no re-recording.
• Test Hygiene: Fixed cross-worker provider leakage; added regression tests for the gate, late registry growth, and schema conversion.

🪟 cua-driver Contract (0.17 → 0.19 verified)

• Timeouts: Dedicated 120s for app.list, 30s for app.* — cutting a slow call short only poisons the serial MCP pipe.
• Routing: launch_app sends correct name/path/bundle_id fields; open_url uses local OS dispatch instead of the driver's ~80s blocking launch_app(urls=…). Verified unchanged on 0.19.

⏳ Pending

• Human TUI recheck per AGENTS.md: /tools desktop group, capability_expand(desktop), click approval, LEAPFLOW_USE_CUA_DRIVER=false reverse check.