fix(sew): validate refined prompts before adopting them in SEWOptimizer

[lamost423]

What happened

While running SEWOptimizer (optimize_mode="prompt", order="zero-order") on a HumanEval subset with deepseek-chat, the optimized workflow regressed from pass@1 = 0.4 to 0.1 after 3 optimization steps.

Inspecting the saved workflow showed the first node's prompt had been replaced by a verbatim echo of the optimizer's own meta-instructions:

Given the above information, please refine the instruction for the 1-th task.
Note that you should always use bracket (e.g. `{input_name}`) to wrap the inputs of the tasks in your refined instruction.
Only output the refined instruction and DON'T include any other text!

The original {question} placeholder was lost, so the node produced garbage for every downstream sample.

Root cause

SimplePromptBreeder.generate_prompt returns the LLM response verbatim, and both _wfg_prompt_optimization_step and _action_graph_prompt_optimization_step adopt it directly:

new_prompt = self._prompt_breeder.generate_prompt(...)
graph_info["tasks"][i]["prompt"] = new_prompt   # no validation

When the LLM echoes the refinement request instead of answering it (models occasionally do this with the nested mutation-prompt + meta-instruction structure), the echo silently becomes the node prompt.

Fix

Add SEWOptimizer._validate_refined_prompt, applied at both adoption sites. A refined prompt is rejected (keeping the original prompt, with a warning) when it:

1. is empty/whitespace,
2. echoes the refinement meta-instructions ("please refine the instruction" / "only output the refined instruction"), or
3. drops an input placeholder (e.g. {question}) that was present in the original prompt — matching the contract the meta-prompt itself states ("always use bracket to wrap the inputs").

Tests

tests/src/optimizers/test_sew_prompt_validation.py — 7 cases covering the real-world echoed response above, empty responses, placeholder loss, legitimate refinements, and the fallback/adoption behavior of _wfg_prompt_optimization_step (via a stubbed prompt breeder, no API calls). All pass; existing test_sew_workflow_scheme.py still passes.

🤖 Generated with Claude Code