Sui MCP: effectiveness eval (Promptfoo) — required acceptance

[bidzyyys]

Sub-issue of #376. The founding "how do we test its effectiveness?" acceptance. Owned by #376 (not shared with #377). Depends on #459 and #460.

Scope

• Promptfoo harness with a corpus of natural-language asks (e.g. "access-controlled treasury with a timelock and a rate limiter") → expected recipe id(s).
• Deterministic scoring vs examples/ as ground truth:
  • tool-call-f1 — correct recipe(s) selected;
  • sui move build on the scaffolded package;
  • r.mvr deps resolve;
  • sui move test passes;
  • zero anti-pattern hits.
• Run pre-release / on schedule (needs Sui CLI + model calls).

Acceptance: harness runs against the live server and reports a pass/fail score per ask; seed corpus authored.