Add TS support for synthetic data, simulation parity, and prompt optimization

[theanuragg]

This PR advances the TypeScript SDK toward the post–July 1st parity goals from #2734 by surfacing synthetic data generation, simulation parity, and prompt optimization capabilities in TS.

Concretely, it:

• Implements a Synthesizer API for synthetic data generation in TypeScript, allowing users to generate goldens from documents, contexts, scratch, or existing goldens, with evolution and filtration configs aligned with the Python implementation.
• Extends the existing simulation surface so that conversation/simulation workflows available in Python are reachable from TypeScript, keeping types and APIs consistent with the current /typescript design.
• Introduces a PromptOptimizer for prompt optimization in TS, providing a metric-driven, iterative optimization loop that mirrors Python’s behavior while feeling natural for TypeScript users.

These changes are wired into the public TS exports and documented, so TS users can now:

• Generate and evolve synthetic datasets in their existing DeepEval flows.
• Use simulation features from TypeScript without dropping back to Python.
• Optimize prompts directly from TS using their evaluation metrics as feedback.

This PR focuses on API completeness and parity; follow‑ups can refine algorithms and add more examples once these surfaces are battle‑tested.

[vercel]

@theanuragg is attempting to deploy a commit to the Confident AI Team on Vercel.

A member of the Team first needs to authorize it.

[penguine-ip]

@theanuragg thanks for the PR. However it is near impossible to review. If you could break this down into scoped changed, I'm happy to take a look.

[theanuragg]

@theanuragg thanks for the PR. However it is near impossible to review. If you could break this down into scoped changed, I'm happy to take a look.

The diff looks a lot bigger than the logical changes because Prettier reformatted a bunch of existing files. The actual code changes are scoped into a few areas:

• New ‎assertTest API: ‎src/evaluate/assert_test.ts adds a standalone ‎assertTest() that calls ‎metric.measure() directly (bypassing the full ‎evaluate() pipeline and Confident AI posting). It returns a typed ‎TestResult and throws an ‎AssertionError with per‑metric details when any metric fails; ‎runMetric() is invoked with ‎showIndicator = false so it doesn’t affect CLI output.

• Prompt usage without API key: ‎src/prompt/index.ts now exposes ‎Prompt.fromText() so you can construct prompts from raw text without requiring a Confident AI API key. The API client is lazily initialized on first ‎pull()/push(), so local evaluation / optimization flows work without remote calls.

• CLI test runner refactor: ‎src/cli/index.ts switches the Jest runner to ‎child_process.spawn with multi‑path binary resolution (tries ‎jest / ‎npx jest on ‎PATH, then common ‎node_modules locations). I also added a ‎test command and cleaned up the console prefixes, but behavior is otherwise the same.

• Synthesizer templates + engine:

▫ ‎src/synthesizer/templates.ts adds Nunjucks‑compiled evolution templates mirroring the Python side (‎EvolutionTemplate, ‎PromptEvolutionTemplate, ‎SynthesizerPromptTemplate).

▫ ‎src/synthesizer/synthesizer.ts wires those templates into a template‑based evolution engine with ‎evolutionMap / ‎promptEvolutionMap, ‎qualifyInput() (with retry), and ‎LLMTestCase generation via structured output. Evolution sampling matches Python’s weighting and properly renormalizes when ‎MultiContext is excluded for no‑context runs.

• Prompt optimizer cleanup: ‎src/optimizer/prompt-optimizer.ts now imports ‎LLMTestCase at the top level instead of ‎require(), constructs prompts via ‎Prompt.fromText() instead of mutating private fields, and uses a small ‎resetMetric() helper that calls ‎metric.reset() when available.

• OpenAI embeddings cost fix: ‎src/models/embedding-models/openai-embedding-model.ts corrects batch cost accounting by splitting total cost across the batch (‎perItemCost = cost != null ? cost / batchSize : null) instead of charging each item the full batch cost.

Did small plumbing updates and minor fixes across the file I worked on.