feat: add FeedbackBuilder utility for structured reflective datasets

[rohitsandadi]

Implements FeedbackBuilder (#264). Addresses #261 (anti-overfit), #262 (task drift), #288 (score visibility in reflective dataset).

Summary

FeedbackBuilder is a configurable utility in gepa.utils that replaces make_reflective_dataset() boilerplate. It replicates the existing side_info extraction logic, then layers on optional enrichments:

• include_score_diff — injects score + gap-from-perfect into each record's Feedback field, so the reflection LM always knows how examples scored (feat: auto-inject score into side_info when not already present #288)
• rationale_field — extracts a named human-rationale field from side_info into a dedicated Rationale record field
• global_constraints — appends constraint text to every record, preventing the reflection LM from copying training data (feat: built-in anti-overfit guard to prevent prompt copying training data #261) or drifting task definitions (feat: task definition preservation — prevent optimizer from changing task parameters #262)

When no FeedbackBuilder is configured, the adapter falls back to existing behavior.

Changed files

• src/gepa/utils/feedback_builder.py — new: FeedbackBuilder dataclass with build() method
• src/gepa/utils/__init__.py — export FeedbackBuilder
• src/gepa/adapters/optimize_anything_adapter/optimize_anything_adapter.py — accept + delegate to FeedbackBuilder
• src/gepa/optimize_anything.py — add feedback_builder to ReflectionConfig, wire to adapter
• tests/test_feedback_builder.py — 13 unit tests

Test plan

• 13 tests covering base construction, each enrichment, combined features, import path, adapter integration
• Full suite passes (383 tests)
• ruff check + pyright clean

[semanticdiff-com]

Review changes with  

Changed Files

File	Status
src/gepa/adapters/optimize_anything_adapter/optimize_anything_adapter.py	5% smaller
src/gepa/optimize_anything.py	0% smaller
src/gepa/utils/__init__.py	0% smaller
src/gepa/utils/feedback_builder.py	0% smaller
tests/test_feedback_builder.py	0% smaller

[ECNUACRush]

Hi! I was looking into #288/#264 and noticed this PR already covers the default Score injection path and part of the FeedbackBuilder direction.

Since this PR has been idle for a while, would maintainers prefer to keep this PR as the main path, or would a smaller follow-up PR focused only on #288's default Score injection be useful?

I’m happy to help with a narrowly scoped PR or tests/docs if that would make this easier to review, and also happy to defer to this PR if it is still the preferred direction.