Add persistent teammate foundation and GLM eval

[ZhijianZhou]

What changed

• add versioned teammate domain models for teams, agents, tasks, and messages
• persist active team, task, and agent state under .clawd/teams/<team-id>/
• restore active team state when ToolContext starts and migrate legacy team metadata
• enforce task state transitions and mark mutating team tools correctly
• improve ToolSearch natural-language matching, wildcard discovery, and returned schemas
• add local tool guidance and provider-safe Anthropic tool-result serialization
• add a repeatable GLM 5.2 teammate evaluation fixture with documented baseline evidence

Why

The existing teammate tools only stored lightweight in-memory metadata, so team
and task state disappeared across restarts. In real GLM 5.2 testing, local
engineering tools were registered but difficult for the model to discover, and
structured tool results were not reliably visible through the Anthropic-compatible
endpoint.

This change establishes the persistent teammate foundation and makes the local
Read/Bash tool path usable before implementing agent spawning and message routing.

Impact

• existing ToolContext.team and ToolContext.tasks callers remain compatible
• team history survives disbanding while the active pointer is removed
• GLM 5.2 can discover and invoke local Read and Bash without falling back to web tools
• the order-discount fixture provides a stable end-to-end target for upcoming teammate runtime work

Validation

• 263 passed via .venv/bin/python -m pytest tests/ -q
• clawd-codex version 0.1.0 (Python)
• authorized z.ai GLM 5.2 smoke test completed ToolSearch -> Read directly
• authorized Bash smoke test reported the fixture's expected 4 passing and 2 failing checks