[marcrdgz]

Performance items 2.3–2.6 of the 0.8.0 plan: fewer per-frame copies in the daemon and client, cheaper link detection, and a quieter idle loop.

Changes

• Daemon frames (2.3)
  • The daemon diffs directly against the pane cache. detach_buffer, which cloned the whole grid every frame, is gone.
  • The previous-frame buffer is updated by copying only the rows that were sent.
  • Each frame is built once per tick and shared by every client as an Arc<DaemonEvent>, instead of being deep-cloned per client.
  • WireCell.text is now CompactString (re-exported from kumo-protocol). Short text is stored inline, so cells no longer each cost a heap allocation.
• Links (2.4)
  • The per-column OSC 8 lookup only runs when ghostty's per-row hyperlink flag is set (new Terminal::row_may_have_hyperlink).
  • The plain-text URL scan is skipped for rows without ://.
• Client (2.5)
  • Grid::apply moves frame rows into the grid instead of cloning them.
  • The blit copies each row as one contiguous buffer slice.
  • Blitting only the touched rows is not possible, because ratatui clears its back buffer on every draw.
  • Removed the write-only Grid::dirty_rows.
• Idle (2.6)
  • New ConfigFingerprint::refresh() re-reads a config file only when its size or modification time changed. Files modified in the last 3 s are always re-read, so a same-length rewrite on a filesystem with coarse timestamps is still caught.
  • Idle park goes from 10 → 50 ms. New connections are accepted on their own thread, which wakes the loop, so CLI/attach latency doesn't regress.
  • A client that is behind and waiting for a retry keeps the loop on the 2 ms cadence.
• Hardened daemon_is_command_driven so it ignores a ConfigReloaded event. That event can fire when concurrent tests change the config env, and the failure was intermittent.

Wire protocol

Byte-identical, with a test comparing against the old String cell layout. The only change is in the Rust API: WireCell.text is a CompactString, so construct it with "x".into().

Benchmarks

Release builds, A/B against the pre-change base, 3 interleaved rounds on a loaded machine. The new benchmark is benchmark_pane_frame_pipeline_120x40.

	before	after
Frame work in daemon loop, full 120×40	190–260 µs	50–66 µs
Frame work in daemon loop, one-row diff	66–87 µs	24–33 µs
Link scans per full frame	+75–100 µs	+6–9 µs
Pane write → frame, end-to-end mean	592–671 µs	485–567 µs

Risks

• Per-client encoding in the writer threads measured about 5–10% slower; this could be noise. Total CPU per frame is still lower.
• Child exits and signals can now take up to 50 ms to be noticed (was 10 ms).
• A config file rewritten after it settled, with the same size and the same modification time, is missed. A normal edit always changes the modification time.

Test plan

• cargo build --workspace && cargo test --workspace && cargo clippy --workspace (run twice)
• Manual: vim/htop, seq 1 100000, resizes and splits while output is scrolling, two viewers on one session, wide characters/emoji, OSC 8 and plain URLs with the link modifier held, selection, config hot reload, kumo session list latency, idle CPU