perf: reduce idle GUI and daemon CPU

[neumie]

Summary

• Present parsed output following terminal input immediately, then coalesce unrelated sustained terminal-only UI updates into one app-wide 20 ms (~50 FPS) cadence while keeping every open window live.
• Keep terminal parsing/input, notifications, key-window focus reporting, and OSC clipboard behavior immediate.
• Tier daemon Git polling by project relevance so visible/subscribed repositories remain responsive without continuously scanning every hidden repository.

Changes

GUI activity rendering

• Batch exact dirty terminal IDs across the application rather than letting independent streams drive separate presentation clocks.
• Use a leading edge: the first output after idle is presented immediately.
• Track a causal output epoch for user input so parsed post-input output bypasses an already-active batch; pre-input backlog cannot consume the priority request, while unrelated terminal IDs remain queued.
• Expire unanswered input-priority requests and do not arm them for empty input.
• Fan each frame out to every rendered pane copy and window sidebar, including visible non-key windows on other monitors.
• Do not infer visibility from key-window activation; focus reporting and notification suppression still follow the actual key window.
• Preserve immediate terminal parsing/input dispatch, bells/notifications, and OSC 52 clipboard reads and writes.

Daemon Git polling

• Keep visible and remotely subscribed projects on 250 ms HEAD and 5 s full-status cadences.
• Move hidden, unsubscribed projects to 2 s HEAD and 30 s full-status fallback cadences.
• Keep targeted HEAD/action refreshes immediate, anchor periodic polling to wall-clock deadlines, and retain cached status for projects skipped in a cycle.

Measurements

• 31 real repositories, 60 seconds: daemon mean 22.93% → 5.69%; p95 94.44% → 18.67%.
• Original controlled single-terminal GUI curve: 12.25% at 60 Hz and 19.20% at 120 Hz.
• Final release, verified-frontmost synthetic sustained output: GUI mean 17.17% at 60 Hz and 17.71% at 120 Hz. Doubling output frequency increased GUI CPU by only 3.1%, consistent with the ~50 FPS presentation cap.
• Earlier 33 ms/background-gated GUI figures are intentionally not claimed for the final implementation.

Test plan

• cargo test --workspace
• cargo test -p okena-terminal input_repaint
• cargo test -p okena-ui activity_repaint
• cargo test -p okena-app -p okena-views-terminal
• cargo build --release
• cargo fmt --all -- --check
• LSP and pi-lens diagnostics across all changed Rust files
• Controlled before/after CPU sampling with macOS proc_pid_rusage