Kimi Code CLI: run the live end-to-end smoke test

[tractorjuice]

Summary

The arckit-kimi extension shipped in v6.4.0 and the governance/security hooks port merged in #668, but the Kimi Code runtime has never been observed end-to-end — Kimi Code requires a Moonshot login before it loads skills, so nothing has been verified against a live instance. This is the one follow-up split out of #659 so it stays tracked.

What to verify (fix the generator, not the test, on any mismatch)

1. Load — start kimi, /plugins install https://github.com/tractorjuice/arckit-kimi.git, confirm the plugin loads with no kimi.plugin.json errors and architecture-workflow injects at session start.
2. Scaffold — arckit init --ai kimi in a test repo; confirm .arckit/ templates land.
3. Skill — run one /skill:arckit-* (e.g. /skill:arckit-requirements); confirm an ARC artefact is written under projects/.
4. Hooks — trigger a Write under projects/ and confirm a governance hook fires: a bad ARC filename is blocked (adapter → exit 2) and provenance stamping runs. Confirm context injection (session start / prompt) attaches via stdout.
5. MCP — confirm the six servers load (the two keyed ones failing without API keys is expected).

Known open items to confirm or fix during the test

• MCP headers on remote servers is undocumented/unverified for Kimi; the two keyed servers (google-developer-knowledge, datacommons-mcp) may not authenticate.
• /arckit:wardley --tidy-owm references .arckit/hooks/owm-tidy.mjs, which arckit init does not stage (cross-extension packaging gap).

Refs: #659 (closed), #668.