fix(sdk): align script grader results

[christso]

Summary

SDK and script-grader authors can now return the finalized pass, score, reason, and optional checks[] vocabulary directly. The SDK builders, Zod schemas, Vitest/workspace adapters, Python helper example, generated assertion scaffolds, and script-grader docs all present that vocabulary as the public surface.

Core script-grader parsing now accepts the finalized JSON protocol, derives aggregate score/pass from checks when needed, and carries reason/checks through the internal evaluator result. It still bridges checks into the current internal assertion_results shape so the artifact writer can be replaced separately by av-kfik.28.6.

Publish-surface considerations: this intentionally updates the experimental @agentv/sdk result surface instead of documenting stale assertions[]/passed aliases as supported public API. The deprecated CodeGraderResult type name remains as an alias to the new result schema, but the wire/public shape is the finalized vocabulary.

Related: av-kfik.28.4

Validation

• bun run build
• bun test packages/sdk/test/define-script-grader.test.ts packages/sdk/test/workspace-grader.test.ts packages/sdk/test/vitest-workspace-grader.test.ts
• bun test packages/core/test/evaluation/graders/script-grader-plain-text.test.ts packages/core/test/evaluation/script-grader-file-backed.test.ts packages/core/test/evaluation/script-grader-multimodal.test.ts
• bun test packages/core/test/evaluation/graders.test.ts packages/core/test/evaluation/execution-metrics.test.ts
• uv run pytest in examples/features/sdk-python
• bun run lint
• git diff --check
• Smoke: bun --env-file=.env apps/cli/src/cli.ts eval run examples/features/script-grader-sdk/evals/suite.yaml --target local_cli exercised the SDK script grader path; script grader per-grader score was 1.0 with all returned checks passing. The overall eval scored 50% because the separate LLM rubric target did not resolve from the example-local target config.
• Attempted live rerun with a temporary combined targets file and --grader-target openai; the LLM grader reached provider execution but failed after retries with pi-ai call failed: Connection error. Script-grader API dogfood is covered; live LLM rubric dogfood is blocked by provider connectivity in this worktree.

---

[cloudflare-workers-and-pages]

Deploying agentv with    Cloudflare Pages

Latest commit:	e8ddf96
Status:	✅  Deploy successful!
Preview URL:	https://2886af14.agentv.pages.dev
Branch Preview URL:	https://grading-sdk-script-api.agentv.pages.dev

View logs