[amanusk]

Problem

Every frame built a ListItem for every row of the active address tab (Txs, Calls, MetaTxs, Events), each with several linear registry lookups, ~10 format! allocations and a Utc::now(). Ratatui then painted ~55 of them.

Since run_loop redraws after every terminal event, each keystroke paid a full frame, and a pasted address arrives as one key event per character. On a contract with ~52k cached calls and ~760 user labels a frame cost ~170 ms: pressing / lagged, and a 66-char paste took ~11 s and appeared character by character. The main thread had burned ~117 CPU-seconds in 9 minutes while all tokio workers together used ~7.

Change

• ListWindow (src/ui/widgets/list_window.rs) mirrors ratatui's List offset rule for single-line rows (no scroll padding): keep the stored offset unless the selection leaves the page, clamp to the list. Callers build rows for window.range() only; render() draws with the selection rebased to the window and writes the absolute offset/selection back into the persistent ListState, so txs_render_state.offset() consumers (enrichment window) keep working.
• The four tabs in src/ui/views/address_info.rs render through it. Row bodies moved into tx_row / call_row / meta_tx_row unchanged; list chrome deduped into list_block / highlighted_list.
• Gap rows keep their rendered positions via a shared rendered_row() (src/app/views/address_info.rs), which apply_rendered / apply_call_rendered now use instead of a private inverse mapping. New *_rendered_selected_with(&gaps) variants avoid recomputing gap positions twice per frame.

Bench

Release profile, 220x60 TestBackend, real ui::draw, synthetic rows (temporary example, not committed):

Scenario	Before	After
Calls tab, 52k rows, 759 labels	170 ms / frame	0.7 ms / frame
Calls tab, 52k rows, 0 labels	79 ms / frame	0.6 ms / frame
Calls tab, 500 rows, 759 labels	1.9 ms / frame	0.7 ms / frame
Txs tab, 52k rows, 759 labels	—	1.1 ms / frame

Frame cost is now flat in list size. Verified interactively on a large-history contract: search and paste are immediate.

Tests

• Unit tests for ListWindow::new (empty, clamped offset, selection above/below page, tail page) and for rendered_row (interleaving, round-trip with tx_pos_to_rendered).
• cargo test, cargo clippy --all-targets (no new warnings), cargo build --release.

Follow-ups

• perf(registry): address/tx label lookups are linear scans, called several times per rendered row #90 (H): O(1) registry lookups (still paid by the color-map rebuild per WS event and by the block/tx detail views).
• perf(main loop): let terminal input win over response bursts and coalesce buffered key events before redrawing #91 (M): prioritize terminal input in the main-loop select!, coalesce buffered key events.
• perf(address): per-WS-event work on large Calls lists is O(n): full re-sort, linear dedup, and a full SQLite rewrite per enrichment #92 (M): per-WS-event O(n) work on large call lists, including the full SQLite rewrite per enrichment.

🤖 Generated with Claude Code