feat(core): add Promptfoo-compatible transform runtime - #1648

Summary

Beads are the source of truth for scope and follow-up. Review summary for: av-kfik.34, av-kfik.34.1.

This PR implements Promptfoo-compatible transform runtime support for authored AgentV eval YAML:

• default_test.options.transform inheritance with test-level tests[].options.transform override
• tests[].options.transform applied once to provider output before grading/assertions
• tests[].assert[].transform applied per assertion only, with isolation from sibling assertions
• file-based transforms such as file://scripts/xlsx-to-markdown.js
• inline string transforms for simple authored YAML cases
• hard rejection of deprecated postprocess with actionable transform guidance
• raw provider output and transformed judge/assertion input traceability in run artifacts

Promptfoo source evidence checked locally at /home/entity/projects/promptfoo/promptfoo, commit 6bfc5a0c7f16f9c4717ac731d276b578e63d0769:

• src/evaluator.ts transformRunEvalResponse
• src/assertions/index.ts assertion-level transform
• src/util/transform.ts
• src/types/index.ts
• examples/config-transform-file and related transform examples

Out of scope / coordinated follow-up:

• Provider-level transform ordering is preserved where AgentV already provides transformed output to the evaluator path; broader provider/target transform parity remains coordinated under av-kfik.34.3.
• Public preprocessor docs/examples cleanup remains under av-kfik.34.2 / av-kfik.34.4.

Validation

Focused and adjacent tests:

• bun test packages/core/test/evaluation/transform-runtime.test.ts -> 7 pass
• bun test packages/core/test/evaluation/loaders/grader-parser.test.ts -> 112 pass
• bun test packages/core/test/evaluation/orchestrator.test.ts -> 97 pass
• bun test packages/core/test/evaluation/validation/eval-file-schema.test.ts -> 31 pass

Repo gates:

• bun run lint -> pass
• bun run typecheck -> pass
• bun test -> 3161 pass, 4 fail in examples/features/vitest-workspace-grader/graders/welcome-banner.test.ts; failures are ENOENT for /app/page.tsx because those example verifier tests expect a prepared workspace, not this repo root.

Live dogfood:

• Used copied primary-checkout .env plus local OpenAI-compatible proxy at http://127.0.0.1:10531/v1.
• Ran source checkout command with local-openai target and local-openai-grader live LLM grader.
• Eval used tests[].options.transform: file://scripts/xlsx-to-markdown.js to convert live provider spreadsheet-style JSON into markdown before llm-rubric grading.
• Result: PASS, 1/1, mean score 100%, threshold 0.6.
• Private evidence: agentv-private:evidence/av-kfik-34-1-transform-runtime at commit cef5663.

Notes

This PR is intentionally runtime-focused and does not migrate public docs/examples away from preprocessors. Those broader cleanup tasks remain in the Beads called out above.