Reduce TUI frame stalls under tool activity

[edwin-zvs]

Summary

• stop running zarvis tool blocks from invalidating the synthesized PTY parser every elapsed second
• add an elapsed-time budget to notification burst draining so heavy session activity yields back to paint/tick/input sooner
• bound zarvis tool expand/collapse work by rebuilding only the retained suffix and capping inline expanded output
• reuse already-rendered pending text metrics when a new zarvis tool entry flushes live text into structured history
• handle partial pending flushes by reusing the rendered prefix and processing only the newly appended suffix
• keep the TUI's local tool-output preview bounded at ingestion
• remove terminal-TUI tool grouping so expand/collapse and live result updates invalidate only one tool block
• update the tool-call grouping spec to document the TUI terminal-rendering exception

Investigation

The repro points at zarvis tool calls and expand/collapse. User traces showed the remaining visible freezes were inside render:terminal_replay, with two clicked expand/collapse frames spending about 328-330 ms in replay and about 343-345 ms in total draw time.

The bisected regression commit was 23f120d (Group adjacent tool calls in clients). The remaining expand/collapse issue was the grouping unit itself: bootstrap and transcript replay could build large ToolGroup items. Clicking one child output mutated the group item, so the renderer had to re-synthesize and replay all grouped siblings rather than one affected ToolBlock.

The terminal TUI now keeps tool calls as independent top-level blocks. That trades grouped chrome for bounded frame work: clicking or updating one tool block invalidates one block, not a potentially large group. The existing spec now says grouping is still allowed for clients that can keep updates bounded, but terminal-style renderers should prefer independently invalidated blocks.

Other fixes in this PR address earlier stalls found while narrowing the issue:

• Running tool signatures no longer include elapsed seconds, avoiding once-per-second parser invalidations.
• feed_task_start recognizes pending PTY text already rendered by the previous frame. If pending text was flushed into a PtyChunk, the renderer reuses the rendered prefix metrics.
• The pending-flush reuse also handles the partial case where the flushed chunk contains an already-rendered prefix plus an unrendered suffix.
• Expanded output is capped inline, and local TUI tool-result history stores a bounded preview rather than cloning full tool output into the renderer.
• Notification draining has a time budget, so session bursts do not monopolize the render loop.

Tests

• cargo test -p agentd-cli notification_drain_stops_on_count_or_time_budget
• cargo test -p agentd-cli run_loop_notification_drain_uses_time_budget
• cargo test -p agentd-cli run_loop_select_biases_input_ahead_of_notifications
• cargo test -p agentd-cli running_tool_signature_ignores_elapsed_seconds_until_controls_appear
• cargo test -p agentd-cli running_tool_block_shows_keyboard_hints_not_buttons
• cargo test -p agentd-cli replay_full_scales_with_history_size
• cargo test -p agentd-cli replay_full_rows_grow_does_not_reprocess_history
• cargo test -p agentd-cli zarvis_steady_state_render_is_cheap_with_many_items
• cargo test -p agentd-cli expanded_tool_output_is_capped_inline
• cargo test -p agentd-cli tool_result_history_preview_is_bounded
• cargo test -p agentd-cli zarvis_tool_expand_collapse_rebuilds_only_retained_suffix -- --nocapture
• cargo test -p agentd-cli zarvis_tool_start_reuses_already_rendered_pending_text -- --nocapture
• cargo test -p agentd-cli zarvis_tool_start_reuses_partially_rendered_pending_text -- --nocapture
• cargo test -p agentd-cli ungrouped_tool_update_stays_bounded -- --nocapture
• cargo test -p agentd-cli live_tool_start_after_render_stays_append_only
• cargo test -p agentd-cli live_late_tool_use_hydration_after_render_stays_append_only
• cargo test -p agentd-cli adjacent_same_tool_task_starts_stay_separate_blocks
• cargo test -p agentd-cli late_tool_use_hydration_stays_separate_from_previous_block
• cargo test -p agentd-cli many_small_events_stay_fast_as_history_grows -- --nocapture
• cargo test -p agentd-cli pty_render -- --nocapture
• git diff --check
• cargo build

Note: cargo fmt --check --package agentd-cli still reports existing formatting drift in unrelated CLI files, so I did not apply repo-wide formatting churn.