Add dataset construction skill reference

[annabellscha]

Summary

• add a Langfuse dataset construction reference following the repo reference-file pattern
• route dataset creation/get-to-good-dataset requests from the Langfuse skill
• bump Claude and Cursor plugin manifests to 1.4.0

Validation

• git diff --cached --check
• Ruby YAML/frontmatter validation for the dataset reference
• staged reference-link validation
• plugin JSON parse check

Note: this is stacked on add-issue-detection-triage so the PR stays scoped to the dataset skill work.

[claude]

LGTM — straightforward addition of a new Langfuse skill reference file following the existing repo pattern.

What was reviewed: the new dataset-construction.md frontmatter/structure against sibling reference files (issue-detection-triage.md, prompt-migration.md) — consistent naming, metadata, and required_access conventions; the SKILL.md routing line addition; and the plugin.json version bumps in both manifests. Checked the referenced academy URL — flagged as unverified but non-blocking since it's a documentation-only pointer, not executable behavior.

Extended reasoning...

Overview

This PR adds a new reference file (skills/langfuse/references/dataset-construction.md) documenting a collaborative workflow for constructing Langfuse evaluation datasets, adds one routing line to skills/langfuse/SKILL.md pointing to it, and bumps the plugin version in both .claude-plugin/plugin.json and .cursor-plugin/plugin.json from 1.3.0 to 1.4.0. No code paths, APIs, or executable logic are touched — this is purely documentation/skill-guidance content plus a version bump.

Security risks

None. The changes are markdown documentation and JSON version metadata. The new reference file explicitly instructs the workflow to require user confirmation before any live dataset mutation (create/upsert/reshape/upload), which is a sensible safety guardrail rather than a risk, and it does not embed credentials, executable code, or novel tool permissions.

Level of scrutiny

Low. This is a pure documentation/skill-content PR with no changes to allowed-tools, CLI permissions, or logic files. It follows the same reference-file pattern as existing files in the same directory (e.g., issue-detection-triage.md, prompt-migration.md) — matching frontmatter shape, required_access metadata, and workflow/rules structure. The version bump is mechanical and consistent with adding a new skill capability.

Other factors

The bug-hunting pass found no bugs. One candidate issue (an unverified academy URL) was raised and ruled out as non-blocking, since it's a documentation pointer rather than something that affects skill behavior if stale. The PR is small, self-contained, and stacked cleanly on a prior scoped PR per the author's note.