fix(desktop): fetch profiles for reaction actors and thread-reply authors

[tlongwell-block]

Problem

Reaction pill popovers sometimes show a truncated pubkey instead of the reactor's display name (80c5f18b… reacted with :party_blob:) — reported by Wes in #buzz-bugs.

Root cause: reaction (kind-7) actors were never included in the profile fetch set. ChannelScreen builds messageProfilePubkeys from collectMessageAuthorPubkeys() (+ mentions/DM participants/agents/typing), and that collector filters on isTimelineContentEvent() — kind-7 excluded. formatTimelineMessages then renders each reactor via profiles?.[actorPubkey] with a ${pubkey.slice(0,8)}… fallback.

The exclusion is old, but the pre-#1500 client-assembled cache accumulated deep channel history, so nearly every reactor had also authored some message in the soup and got a name by side effect. Post-#1500 the timeline is one server-assembled window page; a reactor with no authored row inside the current page falls through to the fallback. (#1514's delta profile fetch is unrelated — it changes how the set is fetched, not what's in it.)

The thread panel had the same gap one level deeper: useIndependentThreadPanel fetched thread replies via its own useThreadReplies cache, so deep-thread reply authors and their reactors never reached the profile batch either.

Fix

• New collectReactionActorPubkeys(events) (kind-7 only, actor-tag aware, skips deleted reactions, lowercase/dedup) — deliberately separate from collectMessageAuthorPubkeys, whose timeline-content semantics other callers rely on.
• ChannelScreen merges reaction actors into the users-batch profile set, over [...resolvedMessages, ...threadReplyEvents].
• useThreadReplies hoisted into ChannelScreen; threadReplyEvents feeds both the profile pubkey collection and useIndependentThreadPanel — thread reply authors/reactors now resolve from the same profile batch.
• Marginal cost: perf: GIN index for e-tag containment + delta profile fetch (scroll-back ~2.1s/page) #1514's per-pubkey delta cache means new reactors cost one small kind-0 batch, never a re-download of resolved profiles.

Note: a few explanatory comments in ChannelScreen.tsx were trimmed to keep the file under the 1000-line CI gate (at 997 before this change); behavior they described is unchanged.

Validation

• Live staging repro (root-cause confirmation + fix verification): Playwright spec driving the built bundle against the staging relay (sprout-oss.stage.blox.sqprod.co via in-cluster port-forward), targeting Wes's exact message/reaction.
  • main @ 4343c6d (RED): popover "80c5f18b… reacted with :party_blob:"; 8 kind-0 batches / 72 pubkeys requested — reactor's pubkey in none of them.
  • this branch @ 0bd12df (GREEN): popover "tho reacted with :party_blob:"; reactor's kind-0 requested and resolved.
• Unit: collectReactionActorPubkeys coverage in formatTimelineMessages.test.mjs; full desktop unit suite 1630 pass / 0 fail.
• E2E: new bridge-mock reaction-names.spec.ts (smoke project) — reactor with no authored message in the window resolves by name.
• pnpm check + pnpm build green; ChannelScreen still under the size gate at 997 lines.

Diagnosis thread: #buzz-bugs 40260a3e. Implementation by Max; root-cause + staging verification by Eva.