perf: Reduce AI TUI rendering overhead for long conversations (#3447)

Fixes keystroke lag in Atuin AI that scales with conversation length.
After extended use (many turns, lots of tool calls with output
viewports), pressing a key had noticeable delay before the letter
appeared.

Three layers of optimization:

- **Skip `sync_view_state` for `InputUpdated`** — every keystroke was
cloning all events, tools, and archived data even though no FSM state
changed. Uses `handle.update_tracked()` (eye_declare 0.5) to skip
rebuilds when values haven't actually changed.

- **Pre-compute turns and `has_command` on the driver thread** — the
view function was rebuilding the full turn structure from raw events and
scanning for `suggest_command` tool calls 3× per render frame. Now
computed once per FSM state change and cached in ViewState.

- **Commit-based element tree pruning** — turns that scroll into
terminal scrollback are tracked via `on_commit` and filtered from the
element tree, keeping rendering work proportional to visible content.
Turn views are now direct children of the root VStack (not nested inside
AtuinAi) so `detect_committed` can see them.

1 files changed, 45 insertions, 3 deletions

diff --git a/crates/atuin-ai/src/commands/inline.rs b/crates/atuin-ai/src/commands/inline.rs
index 70f26c65..989b95c0 100644
--- a/crates/atuin-ai/src/commands/inline.rs
+++ b/crates/atuin-ai/src/commands/inline.rs
@@ -292,6 +292,17 @@ async fn run_inline_tui(
         .bracketed_paste(true)
         .with_context(tui_tx)
         .extra_newlines_at_exit(1)
+        .on_commit(|committed, state| {
+            if let Some(key) = &committed.key
+                && let Some(id_str) = key.strip_prefix("turn-")
+                && let Ok(id) = id_str.parse::<usize>()
+            {
+                let new_count = id + 1;
+                if new_count > state.committed_turn_count {
+                    state.committed_turn_count = new_count;
+                }
+            }
+        })
         .build()?;
 
     // ─── Driver loop ────────────────────────────────────────────
@@ -349,17 +360,48 @@ fn build_view_state(
         skill_names.insert(skill.name.clone());
     }
 
+    let tools = fsm.ctx.tools.clone();
+    let visible_events = fsm.ctx.events[safe_start..].to_vec();
+    let archived_events = fsm.ctx.archived_events.clone();
+
+    let mut archived_builder = crate::tui::view::turn::TurnBuilder::new(&tools);
+    for event in &archived_events {
+        archived_builder.add_event(event);
+    }
+    let archived_turns = archived_builder.build();
+    let archived_turn_count = archived_turns.len();
+
+    let mut visible_builder =
+        crate::tui::view::turn::TurnBuilder::new_starting_at(&tools, archived_turn_count);
+    for event in &visible_events {
+        visible_builder.add_event(event);
+    }
+    let visible_turns = visible_builder.build();
+
+    let mut turns = archived_turns;
+    turns.extend(visible_turns);
+
+    let has_command = visible_events.iter().any(|e| {
+        matches!(e, ConversationEvent::ToolCall { name, input, .. }
+            if name == "suggest_command"
+                && input.get("command").and_then(|v| v.as_str()).is_some())
+    });
+
     ViewState {
         agent_state: fsm.state.clone(),
-        visible_events: fsm.ctx.events[safe_start..].to_vec(),
+        visible_events,
         all_events: fsm.ctx.events.clone(),
         session_id: fsm.ctx.session_id.clone(),
-        tools: fsm.ctx.tools.clone(),
+        tools,
         current_response: fsm.ctx.current_response.clone(),
         is_resumed: fsm.ctx.is_resumed,
         last_event_time: fsm.ctx.last_event_time,
         in_git_project,
-        archived_events: fsm.ctx.archived_events.clone(),
+        archived_events,
+        turns,
+        has_command,
+        committed_turn_count: 0,
+        archived_turn_count,
         is_input_blank: true,
         slash_command_input: None,
         slash_command_search_results: Vec::new(),