fix: exclude unmeasurable LangSmith runs from outliers

[Gautam Sharma (GautamSharma99)]

Summary

Fixes #544.

LangSmith sampling used latencyMs(run) ?? -1 while calculating and filling the latency-outlier slice. When there were fewer measurable runs than the outlier budget, incomplete roots were labeled as slow despite having no latency evidence.

Changes

• Build the outlier candidate pool from non-error roots with a valid, non-negative measured latency only.
• Calculate the proportional quarter cap from that measurable pool, and select only those candidates.
• Continue using the original non-error order to backfill remaining capacity as baseline, so incomplete runs remain available as context without unsupported outlier labels.
• Add regressions for all-missing latency and mixed measurable/unmeasurable pools.
• Add a patch changeset.

Behavior

A run with no usable start/end timestamps can now be selected only as baseline (or omitted by the overall cap), never as a latency outlier. Every root in the outlier bucket has a defined measured latency.

Validation

• pnpm run lint:check
• pnpm run typecheck
• pnpm exec prettier --check (changed files)
• pnpm test — 67 files, 793 tests passed

[changeset-bot]

🦋 Changeset detected

Latest commit: d5a1d3c

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
openwiki	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR