feat(skillet): Skill generated by skillet's new agent-orchestrator pipeline

[gricha]

This PR adds the skillet skill — the meta-skill that routes a
user to the right skillet CLI subcommand — produced clean-room
by skillet's new bundled-agent pipeline.

This PR is just the skill artifact, isolated for review. The
pipeline that produced it (the rewrite that replaces skillet's
multi-phase TypeScript pipeline with a small set of bundled
authoring agents) lives in getsentry/skillet#2.

Generation stats

• Input: spec.yaml (12 behaviors/must_nots, hand-curated)
• Pipeline: skill-writer + eval-writer in parallel, then
  skill-validator + evals-validator in parallel
• Wall-clock: 175 seconds end-to-end
• Tool calls: 6 (skill-writer) + 21 (eval-writer)
• Validators: both ok=true, 0 findings on first pass — no
  re-passes needed

Files

• SKILL.md (118 lines) — router for skillet CLI commands;
  imperative voice, decision table for command selection,
  Don't section for must_nots
• spec.yaml — source of truth (skillet-managed; this is what
  the agents read)
• evals/_judges.ts — 13 canonical named judges
• evals/<id>.eval.ts — one file per spec entry, 14 cases
  total. Uses upstream vitest-evals + skillet's harness via
  @sentry/skillet/evals

Eval results against the produced SKILL.md

./dist/cli.js eval skills/skillet:

• First run: 14/14 passed
• Second run: 13/14 (one judge variance — UsesScopedPackageJudge
  graded a clarifying-questions response 0.0 because the scoped
  package mention came at the end, not the start; the agent
  technically did the right thing, the judge was overstrict)

Known shape vs. what skillet's legacy pipeline produced

This branch's diff vs. the legacy-produced version (on
getsentry/skillet main) is ~263 insertions / ~286 deletions.
Tighter prose, same describeEval ids, same case shapes.

Two cases lost workspace fixtures in favor of judge-only
assertions — a slight regression to address by tightening
agents/eval-writer/references/eval-contract.md in the
skillet repo.

Reviewing this PR

You're looking at the skill artifact end-state. To replicate
the generation:

git clone https://github.com/getsentry/skillet
cd skillet
git checkout experimental/agent-orchestration
npm install && npm run build
rm -rf skills/skillet/{SKILL.md,evals}
./dist/cli.js improve skills/skillet

That regenerates the same files (modulo agent variance).