[jiunbae]

This replaces #226, which I closed by mistake. That PR carried all of these fixes at once; they are now split into focused PRs (#233 popup paint, #234 popup background repaint, #235 destroy-unattached after switch; a renderer perf PR follows once #233 lands).

Problem

An updating display-popup is expensive and flickers. Three things combine:

1. Each popup refresh first painted a rectangle of spaces over the whole popup, then painted the content rows over it. A terminal reading the stream incrementally (a PTY, SSH) can show that blank surface as a flash.
2. Each refresh sent the full popup as a persistent overlay frame. A one-character clock or spinner update in a large popup resent every row. This is also the frame flood behind display-popup: large popup paint floods attach control backlog and detaches the client #216.
3. On Unix the popup reader refreshed after every 8 KiB PTY read. A TUI frame that arrives in several writes was rendered once per read, and the intermediate states were sent too.

Change

Three commits, one for each cause:

• fix(renderer): avoid blanking popups before painting content. Content rows are painted in their final state. Each surface row is padded after its text with the popup style, through the existing ANSI- and grapheme-aware truncator, so shorter and missing rows still erase old cells.
• fix(renderer): send only changed popup rows on child-output refreshes. Popup reader refreshes are marked row-diffable. At the attach output boundary, after the generation checks, the frame is compared with the last full persistent frame actually emitted, and only changed rows are written. Identical frames write nothing. Only frames made entirely of self-contained rows at one column, in strictly increasing order, qualify. Anything else falls back to the full frame: moved or resized rows, nested menus, transient messages. Explicit client refreshes are never diffed. The cache always keeps the full frame, so refresh-client, resize, queue replacement and restore never depend on a chain of deltas.
• fix(popup): batch Unix popup PTY reads into one refresh. After the first byte arrives, the reader waits a fixed 8 ms and drains up to 256 KiB before one refresh. An idle popup does not wake, and sustained output cannot extend the deadline. Windows keeps its existing reader.

Related to #216. This makes far fewer and much smaller overlay frames, but it does not change how the attach control backlog is charged or pruned, so I am not claiming it closes that issue.

Bordered popups (border edges use a different record form) still take the full-frame path here. A follow-up PR stacked on this one handles them.

Tests

• renderer::overlay tests: no all-blank prepaint, each row positioned once, and a screen-model test that short Unicode/combining rows and missing rows erase old text.
• pane_io::persistent_overlay::popup_delta_tests: replaying the delta over the previous frame gives the same screen as the full new frame, including shortened coloured wide-character rows and emptied rows. Identical frames give an empty delta, but the cache keeps the full frame. Restore, hidden overlay, moved rows, resized frames, overlapping drawing and explicit refreshes all fall back.
• popup_reader_batches_split_output_without_waiting_for_eof (Unix only): two writes 1 ms apart come back as one batch, and an idle reader does not return.

Earlier Linux measurement replaying one recorded dashboard popup stream (140x40 client, borderless 99%-height popup, 6 s samples). It compares a build that already had the first commit against one with all three, so it measures the row diff and batching only. Terminal output went from 137,717 B / 663 row draws to 752 B / 4 with status off, and from 138,350 B to 1,385 B with status on. Daemon CPU differences were within one accounting tick.

Platforms tested

Windows 11 (x86_64, native, MSVC, toolchain from rust-toolchain.toml):

• cargo build --workspace, cargo clippy --workspace --all-targets -- -D warnings and cargo fmt --all -- --check pass.
• The tests added or changed here pass.
• The full cargo test --workspace --all-targets is not green on Windows for unmodified main either (on my machine main itself fails a large set of real-daemon integration tests, and the rmux-server lib tests overflow their stack unless RUST_MIN_STACK is raised), so I did not use it as a gate here.

The Unix-only batching code and its test were not built or run on this machine; the Linux numbers above come from earlier work on this change.

🤖 Generated with Claude Code