fix(teleprompt): handle metric_threshold=0.0 in BootstrapFewShot

[udbhav-shrinet]

Summary

Fixes #10603

Description

In BootstrapFewShot, setting metric_threshold=0.0 previously evaluated as falsy in if self.metric_threshold:, skipping numeric threshold comparison and treating the metric result as a boolean. This caused:

1. Valid scores of 0.0 to be rejected (when they should be accepted under >= 0.0).
2. Negative scores like -0.1 to be accepted (because -0.1 is truthy in Python).

Changes

• Updated dspy/teleprompt/bootstrap.py to check if self.metric_threshold is not None:.
• Added unit tests in tests/teleprompt/test_bootstrap.py verifying behavior when metric_threshold=0.0.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] No identified blocking issues. The change is safe to merge based on the reviewed code.

Summary

BootstrapFewShot now checks numeric thresholds whenever metric_threshold is not None, including 0.0.

• Tests cover accepting zero scores and rejecting negative scores at a zero threshold.
• Added checks cover None and positive thresholds.
• No actionable issues found. Tests were not run.

_{Reviews (2) · Last reviewed commit: "test(teleprompt): address review feedbac..." · Reviewed by Greptile}

[VANDRANKI]

This is a community review. It does not clear the merge gate.

I read the diff and ran it. The fix is correct. I would approve, but GitHub does not let me approve without repository access, so this is a comment.

The change: in BootstrapFewShot._bootstrap_one_example (dspy/teleprompt/bootstrap.py), if self.metric_threshold: becomes if self.metric_threshold is not None:. With the old check, an explicit metric_threshold=0.0 was treated as unset, so the metric value was used as a bool. That rejects a score of 0.0 and accepts -0.1 (truthy).

What I ran (Windows, Python 3.13, f1260ff as main, PR head):

• I copied this PR's test_bootstrap_zero_metric_threshold onto main. It fails on main with assert 0 == 1 on its first assertion (threshold 0.0 with score 0.0 should keep one demo). It passes on the PR head. I did not see main fail on the second assertion (threshold 0.0 with score -0.1), because the test stops at the first failure.
• tests/teleprompt/test_bootstrap.py on the PR head, together with the tests from the two other open PRs, passed: 35 passed.

Other places: I searched dspy/ for metric_threshold. The other uses (random_search.py, mipro_optimizer_v2.py, teleprompt/utils.py) only pass the value through to BootstrapFewShot. I found no other truthiness check on it, so this one line is the whole fix.

Duplicates: #10604 and #10605 are open with the same one-line source change (the dspy/teleprompt/bootstrap.py hunks are identical in the three diffs). I ran their tests as well. All of them fail on main at the zero-threshold cases (10604: the 0.0, 0.0 and -0.1, 0.0 cases; 10605: its single zero-threshold test) and pass with the fix. #10604 has the broadest coverage: it also covers the None threshold and the nonzero thresholds. #10605 covers the same None and nonzero cases inside one test. Maintainers will probably want to pick one.

Nit on this PR's test: it calls dspy.settings.configure(...), while the other tests in the file use dspy.configure(...). I did not check whether settings.configure is deprecated.

[udbhav-shrinet]

Updated: switched to dspy.configure() to match file conventions and added test coverage for metric_threshold=None and positive thresholds.