Make daily Rig decomposition benchmark resilient to structured-output and typecheck failures

[Copilot]

The benchmark produced empty solutions when free-form SQL was forced through object-schema validation, while standalone skill typechecking failed because required Node and Copilot SDK types were unavailable.

• Output handling

  • Require s.string for free-form solver results, preserving complete SQL and Markdown fences.
  • Reserve object schemas for structured metadata.

• Standalone skill packaging

  • Add @github/copilot-sdk and @types/node dependencies to the skill package.
  • Pin the benchmark workflow to the updated skill revision.

Example:

const singleSolver = agent({
  model: "medium",
  instructions: "...",
  output: s.string,
});

• Fixes [rig-decomposition-bench] Daily rig decomposition benchmark — 2026-08-27 — tie #499