[marcrdgz]

Plan section 2, items 2.1 and 2.2, plus a dirty-row bug found while measuring them.

Changes

• Stop sampling agent status on every render (2.2). render_dirty computed AgentStatus on every frame of an AI pane, but its only caller (ui.rs) discarded it; the 500 ms refresh_agent_statuses tick was already the sole consumer. render_dirty now returns (), so detection timing is unchanged.
• Clear the dirty flags of every redrawn row (bug). vt.clear_dirty walks rows with a peekable cursor and assumes ascending order, but render_dirty passed HashSet order, so most rows were never cleared and every later frame redrew them. That largely cancelled d9ed455 ("skip unchanged rows"). The fix sorts the rows and adds a debug_assert! on the contract.
• Scan composer rows only where needed (2.1).
  • Non-Codex panes now scan only the rows they are about to redraw instead of reading every cell over FFI each frame. A placeholder that appears is always on one of those rows.
  • Finding a placeholder there switches that frame to a full-viewport scan, so the composer is complete on the same frame.
  • The placeholder row is remembered until a full scan proves it is gone.
  • Codex panes and full redraws keep the full scan, so composer behavior is unchanged.

Benchmarks

Release build, --test-threads=1, µs per render, 120x40 viewport with one echoed key per frame. The benchmark now starts from a full screen of text, and an AI-pane variant was added.

	shell pane	AI pane
baseline (aa5c6fe)	328–362	580–1245
+ agent status	315–377	312–422
+ dirty-flag fix	296–323	274–311
+ composer scan	25–27	27
rebased on main (with #23)	22	18

benchmark_echo_render_cost (client only) is unchanged.

Tests

• Regression tests:
  • render_clears_every_redrawn_row
  • codex_placeholder_on_partial_frame_fills_clean_composer_rows
  • codex_placeholder_is_remembered_until_it_disappears
  • agent_status_does_not_depend_on_render
• Each of the first three was checked to fail with its safeguard removed.
• Gate is green after rebasing on main: cargo build, test and clippy --workspace.
• Manually checked shell, vim, Codex, Claude and opencode panes.

Risk

Because of the dirty-flag fix, unchanged rows are now really skipped. Any place where Ghostty fails to mark a changed row dirty would now leave stale content, where the full redraws used to hide it.