fix(optimization): allow train-only GEPA runs

[drewstone]

What changed

• Accept an empty GEPA selection set while still requiring training data.
• Reuse the training set for GEPA candidate comparison and label that fallback explicitly.
• Release the aligned npm/Python packages as 0.126.4.

Why

The bridge already implemented a train-set fallback, but input validation made that path unreachable. Single-source optimization therefore required a fake duplicate selection case.

Verification

• TypeScript: 3,376 tests passed; typecheck, build, lint, and package verification passed.
• Python: full GEPA + SkillOpt suite passed; one DSPy-only test skipped because that environment conflicts with the pinned GEPA environment.
• Python source distribution and wheel built successfully.

[tangletools]

✅ Auto-approved drewstone PR — 76aa5756

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-24T18:12:47Z}