Add Qwen Code external agent

[suhaanthayyil]

Summary

Adds official Qwen Code support via a native Go external-agent binary, entire-agent-qwen.

Closes #20.

What changed

• Adds agents/entire-agent-qwen with Qwen Code protocol support.
• Installs Qwen command hooks in .qwen/settings.json for lifecycle, tool, compaction, notification, permission, and subagent events.
• Maps Qwen lifecycle hooks into Entire normalized events and records tool/subagent/permission metadata into an Entire-owned sidecar transcript.
• Preserves existing and future Qwen hook settings during install/uninstall, including HTTP hooks, env, statusMessage, allowedEnvVars, custom matcher fields, and unknown keys.
• Supports observed Qwen payload aliases: prompt/user_prompt, tool_input/inputs/input, and tool_response/response/tool_result.
• Handles Qwen Code's empty continuation UserPromptSubmit events without resetting Entire's pre-prompt checkpoint snapshot.
• Extracts modified files from common file path keys, nested edit payloads, and shell write commands.
• Stores stable sidecar transcripts in a repo-scoped OS temp directory and writes .entire/tmp/<session_id>.json markers for Entire session discovery.
• Adds transcript analysis and compact transcript support.
• Adds gated Qwen lifecycle e2e adapter behind QWEN_E2E=1 E2E_AGENT=qwen.
• Updates docs and available-agent tables.

Validation

• mise run build agents/entire-agent-qwen
• mise run test agents/entire-agent-qwen
• mise exec -- external-agents-tests verify agents/entire-agent-qwen/entire-agent-qwen
• mise exec -- golangci-lint run in agents/entire-agent-qwen
• Real local Qwen Code e2e with Qwen Code 0.16.1 + Ollama qwen3:8b through the OpenAI-compatible endpoint: Qwen emitted structured write_file tool calls, Qwen Code created hello.txt and qwen_agentic_report.md, Entire hooks fired, commit condensed metadata, and entire checkpoint list --session 5444285a-6225-4e79-8403-99bb7784b3d3 reported checkpoints 1.

Notes

• This integration targets the Qwen Code CLI (qwen), not the Qwen-Agent Python framework and not Claude Code.
• Local-model success depends on Qwen Code receiving executable structured tool calls from the selected model/backend. The real validation above used Ollama qwen3:8b because it executed Qwen Code tools successfully on this machine.

[suhaanthayyil]

Validation update after real local Qwen test:

• Fixed an observed real-run issue where Qwen Code emits a second empty UserPromptSubmit after tool execution. The adapter now records that payload in the sidecar but does not turn it into a new TurnStart, so Entire's pre-prompt file snapshot is not reset after files are already written.
• Real end-to-end test passed with Qwen Code 0.16.1 + local Ollama qwen3:8b through the OpenAI-compatible endpoint. Qwen emitted structured write_file tool calls, Qwen Code created hello.txt and qwen_agentic_report.md, Entire hooks fired, git commit condensed metadata, and entire checkpoint list showed checkpoints 1 for session 5444285a-6225-4e79-8403-99bb7784b3d3.
• Validation commands passed: mise run build agents/entire-agent-qwen, mise run test agents/entire-agent-qwen, mise exec -- external-agents-tests verify agents/entire-agent-qwen/entire-agent-qwen, and golangci-lint run for the Qwen agent.

One local-only note: repo-level mise run test:ci failed because my working tree has unrelated untracked LangGraph/CrewAI agent directories from a separate branch; the Qwen agent itself is clean and those directories are not part of this PR.