feat(eval): add agent-eval harness for gqlkit vs codegen/Pothos

[izumin5210]

Why

§9 of the TSKaigi 2026 talk needs empirical evidence on whether an AI coding
agent can write idiomatic gqlkit (and naturally avoid N+1 / hide sensitive
fields) without explicit hints in the prompt — and how much the bundled
gqlkit-guide skill helps once it ships.

This PR builds the harness that produces that evidence and captures one
40-run pass against claude-sonnet-4-6.

Summary

New @gqlkit-ts/eval workspace package wrapping [@vercel/agent-eval]:

• 8 evals: {01-crud, 02-relation} × {codegen, Pothos, gqlkit-plain, gqlkit-skill}.
• PROMPTs deliberately omit hints about mappers / objectRef / DataLoader / N+1.
  EVAL.ts checks those properties anyway by inspecting the agent's output.
• shared/ holds the source-of-truth backing types + DataLoader stubs;
  sync-fixtures.ts mirrors them into every eval and copies the
  gqlkit-guide skill (the output of pnpm gqlkit docs) into the
  *-gqlkit-skill variants.
• experiments/sonnet.ts targets claude-code × claude-sonnet-4-6.
• scripts/summarize.ts walks results/ and emits a markdown report.
• REPORT.md captures the latest 40-run result.

Initial findings (see REPORT.md for the full breakdown):

	gqlkit-plain	gqlkit-skill	delta
01-crud mean turns	11.0	7.8	-29%
01-crud mean duration	592s	345s	-42%
02-relation mean turns	11.2	7.8	-30%
02-relation mean duration	466s	387s	-17%

DataLoader / User.email-exclusion assertions pass in every run across
all 40, with no prompt-side hints.

Notes for reviewers

• eval/ is added as a new workspace package (@gqlkit-ts/eval) via
  pnpm-workspace.yaml. It is private: true and not part of the public
  surface.
• biome.jsonc: **/testdata and **/*.md are now excluded.
  • testdata/ was already partially excluded (__generated__ only). The
    new exclusion covers the rest; lint errors that surfaced under
    packages/cli/src/gen-orchestrator/testdata/ were pre-existing and
    unrelated to this PR.
  • *.md exclusion was needed because Biome 2.4 reports an "unsupported
    feature" for Markdown, and this PR adds many .md files from the
    bundled gqlkit-guide skill snapshot.
• packages/cli/src/shared/{directive-definition-extractor,directive-detector,interface-validator}.ts:
  biome --unsafe auto-fix applied useOptionalChain (3 small rewrites).
  Functionally equivalent.
• examples/with-ai-sdk/src/scenario.test.ts: one // biome-ignore line for
  noExplicitAny on a LanguageModelV3 mock that needs any[] to bypass
  strict stream-part typing.
• eval/shared/skill/.claude/skills/gqlkit-guide/references is committed as
  a symlink pointing into packages/cli/docs/ (matches what
  gqlkit docs --claude emits). The sandbox-side copies are materialized as
  real directories via sync-fixtures.ts, so this only matters on the host.

How it ran

cd eval
cp .env.example .env  # ANTHROPIC_API_KEY + DOCKER_HOST for Colima
pnpm sync             # distribute shared fixtures + skill into each eval
for i in 1 2 3 4 5; do pnpm sonnet --force; done   # 5 × 8 = 40 runs
pnpm tsx scripts/summarize.ts 5                    # regenerate REPORT.md