feat: flag "date back from" as blend of "date from" and "date back to"

[ghost]

Summary

Adds a many-to-many phrase set correction for "date back from", which is a nonstandard blend of two correct phrases:

• "date from" (originate in a time period)
• "date back to" (extend back to a time period)

Examples

Input	Suggestions
"issues that date back from 4 years ago"	"date from" or "date back to"
"this code dates back from 2014"	"dates from" or "dates back to"

All inflections covered: date/dates/dated/dating back from.

Test plan

• cargo test -p harper-core --lib -- phrase_set_corrections::tests::corrects_date - 2 true positives pass
• cargo test -p harper-core --lib -- phrase_set_corrections::tests::allows_date - 1 true negative passes
• Full test suite passes with no snapshot changes

Closes #2864

[elijah-potter]

It looks like you need to run just format.

[hippietrail]

Have a look at assert_good_and_bad_results() - it is designed to check linters that can produce multiple suggestions (or might potentially produce bad ones) - you can include both "There are too many open issues that date from 4 years ago." and "There are too many open issues that date back to 4 years ago." - for these you would leave the bad array empty. But you'd want to use them judiciously since sometimes one seems to be much more appropriate than the other for reasons that are hard to pin down when you're a native speaker but can feel intuitively.

So yeah if both sound good to you then use assert_good_and_bad_results but if only one sounds good then just stick with assert_suggestion_result for that one.

[ghost]

Good call, I updated both tests to use assert_good_and_bad_suggestions() with both corrections as good suggestions. For these two sentences both "date from" and "date back to" sound equally natural to me so it made sense to verify both are offered.

[hippietrail]

I see that I listed a good bunch of potential false positives in the original issue. It's probably a good idea to test against all of them unless some are too close to identical.

I also noticed you included all four verb inflections but both the original issue and your tests only cover two of them. You should hunt for real-world examples of the two other tenses to make sure they're actually real mistakes people really make, and not more likely to be false positives.

I recommend trying to do full-text searches on GitHub via its own search interface or using Google if you can. Real-word examples are a lot more informative than making up sentences.

[ghost]

I removed 'dated back from' and 'dating back from' since I couldn't find real-world examples of those being actual mistakes (only the present tense forms appear in the issue examples). Kept just 'date back from' and 'dates back from'.

I also tested the false positives from your issue ('the date back from a client', 'get dates back from Mongoid') and confirmed they do trigger. Unfortunately phrase_set can't distinguish noun 'date' from verb 'date' here. I documented this as a known limitation. This might need a POS-aware approach if the false positive rate is too high.

[elijah-potter]

This code looks good to me! Thanks!