feat: upgrade core agents to GLM-5 with prompt tuning

[pdurlej]

Summary

• Upgrade core agent templates from zai-coding-plan/glm-4.7 to zai-coding-plan/glm-5 in _beagle_v2.md, _arch.md, _coder.md, and _writer.md.
• Tune prompts for GLM-5 behavior: stronger reasoning defaults, clearer constraints, and better tool usage guidance.
• Simplify over-constrained flows where GLM-5 already performs well, while adding explicit quality and guardrail sections.

Why

• GLM-5 has stronger long-horizon reasoning and tool-calling behavior than 4.7, so older prompts can be unnecessarily rigid or noisy.
• This update keeps the same agent roles while aligning instruction style to improve reliability and output quality.

File-by-file highlights

• _beagle_v2.md: move to GLM-5, improve source triangulation/freshness/citation format, keep concise research workflow.
• _arch.md: move to GLM-5, add explicit constraints and permissions, fix section numbering, make deployment strategy selection explicit.
• _coder.md: move to GLM-5, reduce prompt bloat, add tool-first verification workflow and hard safety guardrails.
• _writer.md: move to GLM-5, preserve narrative phases, add anti-slop patterns and length discipline, allow read permission.

Suggestions for maintainers

1. If you want minimal-risk rollout, merge model upgrades first, then prompt-shape changes in a second PR.
2. Consider documenting preferred GLM-5 temperatures per agent in README to reduce drift for contributors.
3. If _beagle_v2.md strict pipeline should be preserved, I can split a compatibility variant (beagle_v2_strict) instead of replacing the current behavior.