Fix APP-4853: restore subagent chips and agent names after session reopen

[seemeroland]

Description

Fixes APP-4853: subagent chips are missing and "Unknown agent" is displayed when reopening a session that used orchestrated child agents.

Root Cause

The TUI startup speedup (PR #13449 / commit 0ac6f59) changed the app to load agent_conversations metadata-only at startup — agent_conversation.tasks is always empty. The initialize_historical_conversations eager child-hydration path was updated to call load_conversation_from_db(&conversation_id) when tasks.is_empty(), but without a fallback: if the read-only DB connection cannot be established (or any read error occurs), the call returns None and the child is never inserted into conversations_by_id.

That missing insertion caused two visible regressions after an app restart:

1. Missing pill bar chips: OrchestrationPillBar::pill_specs filters descendant IDs through history.conversation(&id), so unhydrated children produce an empty pill list → no chips.
2. "Unknown agent" in transcripts: participant_for_agent_id resolves the agent ID to a conversation, then looks it up in conversations_by_id. When that lookup returns None, it falls back to OrchestrationParticipant::unknown_child().

The pre-regression code (QUALITY-768 / #11722) always called convert_persisted_conversation_to_ai_conversation_with_metadata(agent_conversation.clone()) directly, which synthesises a minimal AIConversation from conversation_data (task_id/run_id, agent_name, parent linkage) even when tasks is empty. The TUI speedup PR replaced this direct call with a DB load and dropped the fallback.

Fix

In conversation_loader.rs, when agent_conversation.tasks.is_empty(), try load_conversation_from_db first (to get full task data for rich display), and fall back to convert_persisted_conversation_to_ai_conversation_with_metadata if the DB load fails. The fallback synthesises a minimal AIConversation carrying the correct task_id, agent_name, and parent linkage — exactly what the pill bar and name resolution need.

A new regression test …with_empty_tasks exercises this production code path (tasks == []) to pin the fallback behaviour. The existing test test_initialize_historical_conversations_eagerly_hydrates_orchestration_children passed non-empty tasks, so it always took the direct-conversion branch and missed the regression.

Linked Issue

• Fixes APP-4853

• The linked issue is labeled ready-to-spec or ready-to-implement.

• Where appropriate, screenshots or a short video of the implementation are included below (especially for user-visible or UI changes).

Testing

Added test_initialize_historical_conversations_eagerly_hydrates_orchestration_children_with_empty_tasks which:

• Creates conversations with tasks: vec![] (the production startup shape after TUI speedup)
• Verifies the child ends up in conversations_by_id via the fallback
• Verifies agent_name and run_id are correctly restored
• Verifies children_by_parent and agent_id_to_conversation_id indices are populated

cargo check --tests and cargo clippy --tests -- -D warnings both pass.

• I have manually tested my changes locally with ./script/run

Agent Mode

• Warp Agent Mode - This PR was created via Warp's AI Agent Mode

CHANGELOG-BUG-FIX: Fixed subagent chips missing and "Unknown agent" shown when reopening a session that used orchestrated child agents.

Conversation: https://staging.warp.dev/conversation/4e5df828-99f9-412c-bdf9-b73408bf1141
Run: https://oz.staging.warp.dev/runs/019f5cde-754c-7ebb-bd3e-f487495818ca

This PR was generated with Oz.