[FEA] Agent skill for cuVS integration into Open Source vector engines. #2742

cuVS integration into OSS vector engines (libraries, databases, SQL engines) can be time-consuming, involving building an index interchange layer and language wrappers. OSS devs interested in cuVS need a working cuVS-embedded codebase before they can measure the e2e performance benefit. An agent skill can shorten the time to a functional integration so devs can see the value of GPU acceleration early. The functional integration coded up by the agent can later serve as a reference for a production-ready integration.

Agent skills should:

1. Take user input (target OSS library, APIs, perf goal, Hardware)
2. Provide best practices on building first iteration of API and data interchange layer.
3. Measure E2e perf and profile for perf bottlenecks (via nsight tool traces)
4. Include a test plan to evaluate whether integration meets the user's performance & cost goals.
5. Include a remediation plan to fix the integration code (i.e go back to step 2) based on perf bottlenecks if desired users goals are not met.

As an example -- Qdrant is planning to leverage agentic coding to integrate cuVS for GPU accelerated vector index builds: qdrant/qdrant-to-cuvs#1 & https://github.com/qdrant/qdrant-to-cuvs/blob/master/tasks/intro.md

Wanted to start this thread for a high-level discussion before I start building these skills + evaluation plan with a target OSS library.

Open Ended Questions:

1. Should cuVS host this (docs/, examples/)?
2. First target library?
3. Any interop or bindings work in flight that this should build on rather than duplicate?