perf(tui): eliminate streaming render and persistence hotspots

[Blankeos]

Summary

Streaming sessions — especially tool-heavy subagent transcripts — lagged because several O(transcript) costs ran on the UI thread on every refresh, and each frame did avoidable deep copies. This PR removes the hotspots without changing any rendered output or behavior.

Streaming persistence (biggest win)

• HistoryDAO::replace_messages ran a DELETE plus one autocommitted INSERT per message on the UI thread every 250ms streaming snapshot (per concurrently streaming session). Each statement was its own journal rewrite + fsync, so cost grew linearly with transcript length. It now runs in a single transaction with a cached prepared statement.
• The database is switched to WAL with synchronous=NORMAL, turning snapshots into one cheap log append and keeping readers non-blocking.
• The session→persistence message conversion deep-cloned every part's JSON payload on top of the caller's own clone; it now moves the owned parts (one deep copy per snapshot instead of two).

Tool-heavy / subagent transcript layout

• Every streaming layout refresh reformatted all tool parts of the streaming message: JSON clone → serialize → reparse → row formatting → wrapping → syntect diff highlighting, per part. Rendered tool rows are now cached per (message, part), validated by a structural hash of the payload, and scoped to the actively streaming message so memory stays bounded (cache drops when the stream ends or moves to a new message).
• Tools that can never join exploration/task groups (e.g. edit/write, whose args embed whole file contents) no longer build parsed tool info at all on the grouping check, avoiding a deep JSON clone per part per refresh.

Subagent tab switching (ctrl+x down, left/right)

• Render caches are retained for the current session family (root + subagent children), so cycling between subagent tabs is a warm-cache render instead of a full transcript rebuild. Only chats outside the family release their caches, which keeps the memory bound (memory no longer scales with every session visited during a run).
• The status-bar usage text re-walked the whole transcript on every streaming layout refresh, re-serializing tool args (entire file bodies for edit/write) and re-parsing tool payloads to estimate tokens. The completed-message token base is now cached per (session, message count, streaming index); only the streaming message's already-tracked counter changes per refresh.
• Editor-location inference over the streaming message deep-cloned tool args per refresh and allocated three display strings per path candidate per rendered line. It now reads payloads by reference, precomputes mention needles once per pass, and reuses a single line-text buffer.

Per-frame allocations

• Chat::render deep-cloned every visible line (all span strings) each frame, at up to 25fps while streaming. The viewport now borrows lines from the layout cache; highlight/selection helpers were made lifetime-generic with Cow-preserving span splitting so cold paths still work unchanged.
• The event loop cloned the entire terminal cell grid after every full frame to support the isolated subagent spinner fast-path, even when that path could never run. The copy is now kept only while the fast-path is active and reuses its allocation via clone_from; spinner frames restore the base buffer in place instead of cloning it twice.
• isolated_subagent_spinner_interval built the full subagent tab list (session walk + string/theme allocations) every event-loop iteration just to answer "is this a child session"; it now does an O(1) parent-id check. The spinner render resolves its one color via a new allocation-free SessionManager::descendant_position instead of materializing all tabs.

Memory

• Background chats outside the current session family release their rebuildable line caches; Chat::clear now clears the ordered markdown part cache, which previously survived session resets.

Bug fixes found along the way

• editor_location_for_selection iterated start_line..=end_line where an unclamped selection can have end_line == usize::MAX, spinning a core at 100% (this also made cargo test hang on the pre-existing chat_selection_action_bar_shows_before_mouse_up test). The scan is now bounded by the cached location table.

Testing

• New unit tests: tool-row cache hit/invalidation produces identical output and re-renders on payload changes; completed messages don't populate the cache; family-scoped cache retention on session switches; the cached usage base agrees with a fresh walk and invalidates on appends; release_render_caches rebuilds byte-identical output; descendant_position matches descendant_sessions ordering; the selection-location scan terminates on unclamped selections.
• Full suite: 1009 passed. The 4 remaining failures (ollama catalog tests, test_handle_models_shows_ollama_without_connection, test_edit_tool_renders_codex_style_diff_summary) fail identically on main in this environment (hardcoded developer cwd / shared-cache flakiness) and are unrelated.
• Manual smoke test of the TUI in tmux: home screen, command palette, input, sessions dialog, and quit all work; the database opens in WAL mode against a pre-existing schema.
• cargo fmt and cargo clippy are clean relative to main (same warning count, none in touched code).

  

[cloudflare-workers-and-pages]

Deploying crabcode with    Cloudflare Pages

Latest commit:	b475102
Status:	✅  Deploy successful!
Preview URL:	https://875d095d.crabcode.pages.dev
Branch Preview URL:	https://cursor-optimize-tui-renderin.crabcode.pages.dev

View logs