[Bug]: selector-filtered _changes is slower the first time after creating a Mango index

### Version

3.5.1, and `main` branch

### Describe the problem you're encountering

Despite the fact that `_changes` does not use indexes to speed up selector filters, the first time it is requested after creating a Mango index, it is slower, which possibly indicates it's spending time updating the index.

### Expected Behaviour

The time taken for `_changes` to respond should not depend on the existence or otherwise of Mango indexes.

### Steps to Reproduce

To demonstrate this, I created a DB containing a million docs of the form `{ "_id": "doc-N", "value": N }` for N from 1 to 1,000,000. Then I timed how long it took to run this filtered changes request:

```sh
# uses this shell function:
cdb () {
curl -gsu 'admin:admin' "http://127.0.0.1:5984$1" "${@:2}"
}

time ( cdb '/db/_changes?filter=_selector' -X POST -d '{ "selector": { "value": { "$lt": 100 } } }' )
```

This would take around 12 seconds.

Then I created an index covering the filtered field and re-timed:

```sh
cdb '/db/_index' -X POST -d '{ "index": { "fields": ["value"] }, "name": "by-value", "type": "json" }'
time ( cdb '/db/_changes?filter=_selector' -X POST -d '{ "selector": { "value": { "$lt": 100 } } }' )
```

The first time after creating an index, `_changes` takes over 30s to complete. On later requests it's back to around 12s, possibly very slightly slower.

This effect shows up regardless of which fields the index covers; even if it does not cover the `_changes` selector, the `_changes` request is still slow.

### Your Environment

macOS 15, CouchDB 3.5.1, single node, database created with q=2, n=1

### Additional Context

_No response_

Reproduce the slowdown with the provided curl commands using a large database, then compare the first and later selector-filtered _changes requests before and after creating the _index entry. Trace the _changes selector-filtering path alongside index creation or update handling. Done means the first filtered _changes request is not delayed by unrelated Mango indexes, with regression coverage for the reproduced case.

Written by the indexing model from the issue text.