Check every agent integration against a live session #185

The integrations added in #149 were built from each agent's source code (Codex, Gemini CLI, Qwen Code, OpenCode, Kimi Code, Cline) or its docs (Copilot CLI, Cursor, Factory Droid), and checked with recorded payloads and transcripts, but most haven't been run against the real agent. Open questions:

• Copilot CLI: does additionalContext from a userPromptSubmitted command hook reach the model? The docs and the changelog disagree.
• Cursor: does a non-zero Shell exit fire postToolUse or postToolUseFailure? What keys does tool_output have, and which events does the Cursor CLI fire?
• Factory Droid: what shape is the Execute tool_response, and is the hook's session_id the transcript's session id?
• Kilo Code: are the config and data paths the same as OpenCode's under kilo?
• Gemini CLI: what are the exact tool call id and diff stat keys in a real transcript?

Each adapter accepts every plausible shape and never fails a hook, but a wrong guess means that feature silently does nothing for that agent.

Fix idea: for each agent, trodden connect <agent>, do a small task twice, then check trodden status, trodden list or trodden rejected, and the recall on the second run; fix the adapters from what's observed.