Stop conversation-gap expansion from inserting unhydrated grandparents

[Pitchfork-and-Torch]

Summary

ConversationGapAncestorHydrator splices a skipped grandparent tweet id into candidate.ancestors from the parent's in_reply_to_tweet_id, then VF and convo_marshaller consume that list. It never TES-fetched the spliced id.

Consequences on Following today (and on For You if PR 97 lands the same hydrator):

• ancestor_texts was built from the pre-splice id list, so muted-keyword matching missed grandparent text
• ancestor_users was never written, so Brazil / socialgraph mute-block (PR 101) missed the grandparent author
• tombstone_ancestor_ids could not mark a deleted grandparent, so the conversation module still tried to render it

This PR TES-fetches spliced ids, then hydrates texts, users, and tombstones from the expanded list. CoreData ancestor_users are merged, not wiped (TES miss does not clear parent/root authors). One class, one file. No ranking changes.

This is not PR 97 (Phoenix wiring of the hydrator) and not PR 101 (the mute/block filter).

Proof

• Entry: ConversationGapAncestorHydrator.expand_ancestors_for_gap inserts grandparent tweet id
• Sink: VFFollowingCandidateHydrator / convo_marshaller / FollowingViewerMutedKeywordFilter / AuthorSocialgraphFilter / Brazil2026ElectionFilter
• Break: TES was only fetched for original ancestor ids; texts used the pre-splice list; ancestor_users was never set
• Viewer effect: a skipped-conversation grandparent from a muted, blocked, listed, or deleted account still appears in the Following thread module
• Twin: CoreDataCandidateHydrator.build_ancestor_users and the hydrator's own ancestor_texts/tombstone helpers, which already run for the original parent and root

Test plan

• spliced_grandparent_gets_text_users_and_tombstone fails on main, passes on this branch
• deleted_spliced_grandparent_is_tombstone fails on main, passes on this branch
• deep_chain_inserts_grandparent and short-chain keep cases still pass
• batches_unique_parents_across_candidates now expects a second TES batch for spliced ids
• unit tests added; cargo test cannot run in the public dump (no Home Mixer Cargo.toml)