chore: fix sampled field semantics on JudgeResult

[jsonbailey]

Summary

• Inverts the sampled field semantics on JudgeResult: sampled=True now means the evaluation was sampled and run; sampled=False means it was not sampled (bypassed by the sampling rate).
• Previously sampled=True indicated the evaluation was skipped, which was counterintuitive and inconsistent with the field name.
• Adds an early-return guard in LDAIConfigTracker.track_judge_result() to skip tracking when not result.sampled.

Test plan

• Updated existing sampling-rate test to assert sampled=False on skip path
• Added sampled=True assertions to both successful evaluation tests
• All 114 tests pass

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes a public result-field meaning and alters when judge metrics are emitted, which can affect downstream consumers and analytics even though the code changes are small.

Overview
JudgeResult.sampled semantics are inverted so sampled=True now indicates the judge evaluation ran (and remains False when skipped by sampling_rate). Judge.evaluate() sets sampled=True only on the run path, and LDAIConfigTracker.track_judge_result() now early-returns when not sampled to avoid emitting metrics for skipped evaluations.

Tests are updated to assert the new behavior for both success and sampling-skip cases.

^{Reviewed by Cursor Bugbot for commit c4e949e. Bugbot is set up for automated code reviews on this repo. Configure here.}