Tier 3: draft docs/AI_COLLAB.md — AI-driven PR conventions

[Xinze-Li-Moqian]

Part of umbrella #48. Maintainer-alignment task — both maintainers use AI to drive PRs; we need a shared convention.

Context

Both maintainers (@Xinze-Li-Moqian, @poet77) routinely use AI agents to draft PRs / refactors. Without an explicit convention, two AI-driven workflows can:

• Produce overlapping or conflicting refactors
• Drift in commit/PR style
• Generate large PRs that resist human review
• Lose attribution / context for future maintainers

Proposed docs/AI_COLLAB.md content

(Strawman — refine in this issue's comments before drafting the file.)

1. Transparency & scope discipline — PR descriptions state what the human directed (not what AI generated)
2. Plan before code — non-trivial refactors (>200 LOC or ≥3 files) need a design comment in the umbrella issue before any PR opens
3. PR size cap — soft limit ≤ 300 LOC per PR; explicit justification if larger
4. Don't auto-merge each other's AI PRs — even if "looks plausible," reviewer must actually read
5. Branch isolation — each maintainer's WIP stays on their own branches; no cross-branch AI edits
6. Shared lessons — non-trivial conventions go into docs/CONVENTIONS.md or REFACTOR_PLAYBOOK.md, not private AI memory
7. WHY in commit/PR body — diff already shows WHAT; body explains the why-not-alternatives

Acceptance

• docs/AI_COLLAB.md exists with content agreed by both maintainers
• CONTRIBUTING.md cross-references it
• Both maintainers' AI sessions consistently apply the conventions (verified by spot-check on next 2-3 PRs from each)

Notes

Both maintainers assigned. Use this issue thread to converge on the strawman before drafting the doc — this is the kind of thing where unilateral drafting can feel imposed.