fix(mobile): restore ordinary DM behavior for Bestie - #8987

🤖 P2 · Keep older Bestie history reachable after an empty filtered page

When the channel-window query falls back to legacy history, NostrFilters.messages loads the newest 200 mixed-kind events. If those are non-broadcast thread replies (with their root older than the page), this change filters every message out. Previously Bestie displayed those replies in its flat timeline. _MessageList now returns _EmptyMessages; unread backfill skips empty entries, scroll pagination has no list, and status-bar paging rejects zero display entries.

User/system effect: An existing Bestie conversation appears empty and its older root and thread cannot be reached through conversation paging.

Test: Make the channel-window request unavailable, return 200 replies whose root is on the next legacy page, then open Bestie and verify that older history remains reachable.

🤖 Confirmed independently at c6fef5d80 (Carl, an automated reviewer, commenting via Wes's GitHub account). On the legacy fallback, a page of only non-broadcast replies filters to zero rows. message_list.dart:155,829 then skips unread backfill and returns _EmptyMessages, which has no pagination trigger. The Bestie journey seeds the root in its first fake page, so it can't catch this. Fix: keep fetching older pages until a visible root or the real history boundary, and add a reply-only first page to the existing journey.

🤖 P2 · Reserve the working pill’s space when measuring the reply label

When a single agent is still working in a thread with existing replies, the label starts at 61 points. This measurement permits it to use the parent width minus only 24. If its measured width falls between those two limits, 61 + textWidth exceeds the parent constraint: the SizedBox shrinks, but the right-aligned OverflowBox keeps the wider text and ClipRect removes its leading content. Narrow footer space with enlarged accessibility text makes this reachable.

User/system effect: The visible reply count can be clipped off while the agent is working, despite room to wrap it correctly. Text measurement and the animated width need to account for the actual working-state space.

Test: Show an existing single-agent reply summary while that agent continues typing, use a narrow viewport with enlarged text, and verify the leading count remains visible throughout the handoff.

🤖 Confirmed with a throwaway widget test at c6fef5d80 (Carl, an automated reviewer, commenting via Wes's GitHub account). With the agent working and 12 replies at 160px/1x, 200px/2x and 320px/3x, exactly 37px of leading text is clipped: the label is measured at maxWidth - 24 but placed at 61. It also happens without wrapping whenever textWidth > maxWidth - 61. Fix: measure at maxWidth - 61 while existingLabel is set, so width and placement agree.

🤖 P2 · DM agent typing disappears when its thread root isn't loaded

In DMs this now keeps every thread typer out of the main typing row. The only remaining place to show that work is the inline footer, and message_list.dart:1017-1027 renders it only under a root already in the loaded window. Nothing fetches the root when typing arrives.

Failure: reopen Bestie (or any agent DM) while the agent works in an older thread outside the newest loaded roots. There's no typing row and no footer, so the agent looks idle until it replies. On main, the DM-agent clause kept that work visible in the main row.

Fix: keep the main-row fallback for DM thread typers whose root isn't in timelineEntries. Alternatively, fetch or retain the working root. Add an unloaded-root step to the existing DM typing journey; today's typing tests always seed the root.

🤖 P2 · Bold Text breaks the reply label

TextPainter measures with the raw labelMedium style, but Text merges DefaultTextStyle and applies FontWeight.bold when MediaQuery.boldTextOf(context) is set (Flutter widgets/text.dart). The OverflowBox then forces the bolder text into the narrower measured width, so 12 replies wraps or ellipsizes.

Scope: every single-participant footer now takes this path, including ordinary human threads that used _AvatarStack on main. With iOS or Android Bold Text on, that becomes a visible regression across channels, not only during agent work.

Fix: measure with the same effective style Text will render (merge DefaultTextStyle and bold), or use a TextPainter-free layout so the label doesn't get a tight width. The test font is the same width bold and regular, so a widget test needs to assert the resolved style or wrap/ellipsis state, not pixels.

🤖 Follow-up fix: #9013 resolves DefaultTextStyle and the platform Bold Text override before measuring the reply label. A regression using bundled Inter reproduces the original width mismatch and covers Bold Text off/on/off in idle and working footers at 1x/2x scale. Full mobile analysis and tests passed (4,252 Flutter, 4 push configuration, 27 crypto), as did mandatory push hooks. The follow-up is draft pending app/human verification; this thread remains open until the fix lands.

🤖 P3 (optional) · First reaction remounts the reply footer

The reaction row is now wrapped in an unkeyed Padding, the same type as this footer Padding. When a message gains its first reaction (or loses its last), updateChildren matches the reaction Padding to the old footer element. That remounts the footer, and ReplyActivityHandoff loses its controller and refs, so a running handoff snaps. Agents often react on pickup, which makes this a realistic mid-work case. A probe widget test confirmed the state isn't preserved.

Fix: add a ValueKey to the footer Padding, or to the reaction wrapper.