Currently, Wikibase stores semi-structured edit summaries in the text of a revision comment, and then disassembles and properly formats them on display. This conversion is inherently lossy, leading to problems like T186035.
Now that we have a JSON blob per comment available (the `comment_data` field in the `comment` table, thanks to T166732), we should use that instead to store the information in a structured way. The `comment_text` should then, according to the documentation in the SQL file, be that information rendered in the content language (i. e. English on Wikidata).
Currently, you may have a `Summary` with members like the following:
- module name: `wbsetlabel`
- action name: `add`
- language code: `en-gb`
- comment arguments: (empty list)
- summary arguments: “test label”
- user summary: “user-specified summary (API parameter)”
(This corresponds to a `FormatableSummary` with the same fields, except that the module and action name are collapsed into one field, the message key `wbsetlabel-add`.)
These are then combined by the `SummaryFormatter` into a single `comment_text` like the following:
> /* wbsetlabel-add:1|en-gb */ test label, user-specified summary (API parameter)
`AutoCommentFormatter` then picks this apart again and formats it in the user language into a message like the following:
> !!Added [en-gb] label:!! test label, user-specified summary (API parameter)
(The !!highlighted!! part is in reality visually muted, not highlighted, but I can’t replicate that in Remarkup.)
* What implications does this have for external tools? Some of them also format the comment text like Wikibase does (e. g. [WDVD](https://tools.wmflabs.org/wdvd/index.php?lang=en&limit=50&description=on&labels=on&sitelinks=on), which queries the wiki replicas directly, or [EditGroups](https://tools.wmflabs.org/editgroups/), which listens to the [`recentchange` EventStream](https://stream.wikimedia.org/?doc)) – if we change the comment text to be the rendered comment in the content language, they will break.
* Can we make the structured comment data available to AbuseFilters? This could solve e. g. T47252; see also T205254 for more general information.
* Do we want to migrate old comments to `comment_data`, or should we leave them alone and keep `AutoCommentFormatter` so we can still render them?