Merge Dev - Add LangSmith integration and bump cli version (0.3.17)

[aliroberts]

Summary

• Add LangSmith as a pluggable eval backend, enabling Weco to optimize code against LangSmith datasets and evaluators
• Add an interactive browser-based setup wizard that launches when LangSmith args are missing in a TTY session
• Add dataset split support (--langsmith-splits) following LangSmith best practices
• Bump version to 0.3.17, bump minimum Python to 3.9, remove unused fastapi/slowapi deps

LangSmith Integration

Eval Backend (weco/integrations/langsmith/backend.py)

Pluggable backend implementing register_args(), validate_args(), and build_eval_command(). Adds 12 --langsmith-* CLI flags covering dataset, target, evaluators, splits, adapters, dashboard evaluators, and custom metric functions.

Evaluation Bridge (weco/integrations/langsmith/bridge.py)

Subprocess runner that dynamically imports user code, resolves evaluators (custom module:function specs, LangSmith built-ins, or local evaluators.py), runs client.evaluate(), polls for async dashboard evaluator scores, and prints
metrics in Weco's key: value format. Supports target adapters for LangChain runnables and single-input functions.

Setup Wizard (weco/integrations/langsmith/wizard/)

Browser-based single-page app served from a local HTTP server. Guides users through API key setup, dataset selection (with split picker), source file selection, target function discovery (AST-based), and evaluator configuration. Mutates the
CLI args namespace on submit.

Dataset Splits

All layers support --langsmith-splits to filter evaluation to specific dataset splits (e.g. opt, holdout) instead of requiring separate datasets per split. The wizard fetches available splits and shows a chip picker.

CLI Changes (weco/cli.py)

• Added --eval-backend flag with backend registry (_EVAL_BACKENDS, _load_backend())
• Made --source, --eval-command, --metric, --goal non-required when using a backend
• Backend args registered dynamically via backend.register_args()

Example

examples/langsmith-zeph-hr-qa/ — HR QA agent optimized against a LangSmith dataset with custom evaluators, dashboard LLM judges, and a gated metric function. Uses dataset splits for train/holdout separation.