Should flag noun vs verb grammatical number disagreement in relative clauses

For instance: Here are some of the **file types** that **has** built-in support.

"file types" is plural but "has" is singular.

Also wrong would be "Here is a **file type** that **have** built-in support".

Some more:
- You may want to update the **country code and other settings** that **works** best for you.
- I require someone who can suggest **delicious recipes** that **includes** foods which
are nutritionally beneficial
- I want **a system** that **allow** users to register and save their vehicle information
- the flexibilities for users to enable **features** that **requires** TCP connections
- Here you can find the SO100 **docs** which **is** deprecated.
- rather a frequent **occurrence** that **happen** as part of routine work
- 1-2 Technical **interviews** that **resembles** a problem SigNoz solves for
- there is a **script** that **generate** the boilerplate of the new tool
- the ability to create **applications** that **does not** require the user to install the Framework
- A MCP **Server** that **enhance** your IDE with AI-powered assistance
- Managed performance **runners** for GitHub Actions that **provides** 3
- Always free **tools** for web developers that **includes** CSS minify unminify
- Simple **writer** that **rotate** log files automatically based on current date and time
- there are some **differences** that **prevents** them being redundant
- teler-waf is a Go HTTP **middleware** that **provide** teler IDS functionality
- **Package** that **provide** functions for simplified creation of pointers from constants of basic types.
- A **package** that **provide** many helpful methods for working with rest api.
- A **linter** that **detect** the possibility to use variables/constants from the Go standard library.
- A MCP **Server** that **enhance** your IDE with AI-powered assistance for Intlayer i18n / CMS
- to evaluate its own performance on **domains** that **requires** deep expertise

I started work on this and it turns out to have a ton of unexpected false positives, possibly more than true positives in my corpus - and you can see how common the true positives are above. I stopped collecting before I got to the end...

If you want to tackle this, I'll be happy to share my work so far as a draft PR etc.

Review the listed relative-clause examples and the author's draft work first; no repository file, test, or entry point is named. The main unresolved work is defining acceptable false-positive behavior and validating the check against the examples and the broader corpus.

Written by the indexing model from the issue text.