Migrated from spboyer/waza#385

Summary

Design a shared eval and grader registry for waza — inspired by OpenAI Evals' 800+ community registry but adapted for agent evaluation.

Context

From docs/research/waza-vs-openai-evals.md, the registry gap is waza's #1 competitive disadvantage (Row 10). This epic covers the full design.

Sub-issues

• feat: per-turn checkpoint graders (closes #358) #386 — Map OpenAI Evals format to waza graders
• feat: add MCP server mocks (closes #363) #387 — Design Go-module-style grader/eval references
• feat: per-task tool metrics with structured arg matchers (closes #366) #388 — Evaluate registry backend options (Git/OCI/Releases/federated)
• refactor: decouple execution events from Copilot SDK #389 — Design composable eval construction
• feat: Decouple ExecutionResponse from Copilot SDK event types (#10) #390 — Grader plugin extensibility design (WASM/external)

Peter's Ideas (verbatim)

• The registry of shared evals is interesting. Graders particularly.
• OpenAI's are all in their repo as YAML files
• Consuming their format could be interesting
• Go module style: just point to a repo and that is your grader or eval
• Being able to construct your eval from a set of known graders is interesting

Deliverable

Design document at docs/research/waza-eval-registry-design.md — design only, no implementation.

Non-goals (for now)

• Implementation — this is design research only
• NOT a single JSON file for the registry — needs to be more robust
• Not building the actual CLI commands yet