You need to edit a page that has some hardcoded data for it, like the one I created here:

I pushed the ve.RegExpPool proof of concept to gerrit (1287885). At scale, executing one large RegExp seems 33% faster than executing multiple individual RegExps. But right now the savings don't compensate for the extra time overhead of operating the pool logic (though this could likely be optimized).

javascript
const pool = new ve.RegExpPool();
const individual = [];
for ( let i = 0; i < 1000; i++ ) {
      const regExp = new RegExp( `id${ i }`, 'gu' );
      pool.register( regExp );
      individual.push( regExp );
}
pool.buildRegExp();
const bigRegExp = new RegExp( pool.patterns.map(
      ( pattern, i ) => `(?=(?<p${ i }>${ pattern })?)`
).join( '' ), 'gu' );

Ok it may be possible to use optional zero-width lookahead assertions with named captures inside:

We did think about whether the giant disjunction RegExp could use named captures to say which rule matched:

Draft SLO here: https://wikitech.wikimedia.org/wiki/User:DChan_(WMF)/EditSuggestionSLO

Debugging jointly, we can see this in @awight's browsers, even in private/incognito mode, but not in my own. Increasing vertical resolution during the session seems to correct things.

Boldly estimating this as 5 points

In which contexts would we *not* want to suggest replacing [٠١٢٣٤٥٦٧٨٩] with [0123456789]? I imagine:

Oh, this (merged) patch should also have been tagged:

On discussion, we decided it is viable, and most feasible, to launch without the feedback incorporating any automated content at all. The user would then take responsibility for manually including in the feedback message any context that might be useful as a part of their feedback.