[darkoatanasovski]

Linear: AI-1624

Why

The Nora template's instructions assumed every case was a chat message sitting in a channel or thread. To review custom check types, Feeds activities and comments, and user profiles, Nora needs to know what context each of those brings, and what it may quote.

Changes

• instructions.md: a case may carry an entity_type, and the instructions say what context each kind comes with:

  • chat message (or no entity_type, so today's bundles still read the same): its conversation
  • Feeds activity or comment (stream:feeds:...): an activity with its comments and reactions, the case in place, marked by case_id, with its text left out because it is in the case. Reactions show how people received the post, not whether it violates a label
  • user profile (stream:user): the profile's fields as the case text, one field: value per line
  • any other entity type (custom checks): the text on its own, with the entity type as a hint. It is not judged uncertain just because there is no conversation

  The quoting rule now also says that evidence never comes from the activity under a comment, another comment, or a message around the case.

• skills/harassment.md: "who the content addresses" now covers the post a comment answers, not only a conversation.

• Simulations: four new ones next to the chat cases: a product review with no conversation, a Feeds comment attacking the post's author, a Feeds activity that quotes abuse to report it, and a spam user profile.

• README and CHANGELOG updated.

The output contract (internal/moderator/contract.md) does not change. Judgments, the flag assessment and the summaries stay the same for every kind of case.

This sets the bundle shape that the chat-side tickets (AI-1620 custom checks, AI-1621 Feeds, AI-1623 user profiles) should send: entity_type on the case, plus activity instead of conversation for Feeds.