perf(watch): re-sort on push + skip redundant idle redraws

[jiunbae]

Problem

muxa watch felt like it updated "a beat late" (한 박자 늦게), especially with sort = ["state", "latest"].

Root cause: a streaming SingleAgent transition updates the row in place but the loop never re-sorts on a push — ordering only refreshes on the 5 s Full fallback tick (STREAMING_FALLBACK_INTERVAL). When the sort key is a pushed field (state), the badge changed instantly but the row's position lagged up to 5 s.

Separately, the run loop repainted unconditionally every input-poll tick (~62 fps) even on a fully idle UI — wasted CPU.

Changes

1. Re-sort on push — apply_outcome now calls resort_rows_preserving_selection() after applying a SingleAgent. The merge is surgical and sort_by is stable, so only the changed row moves (no "all rows jumped" jitter); selection is pinned by pane_id so the cursor stays on the same agent. State-sorted views now reflect transitions immediately.
2. Dirty-gated rendering — repaint only when something changed (input / refresh outcome / preview recapture) or a 1 s idle cadence elapses. The cadence keeps the Activity column's second-granularity relative_time current without the constant full re-render.

Notes / ruled out

• Reconciler (30 s) only handles closed-pane reaping / codex caps / stuck detection; snapshot debounce (200 ms) is disk-only; streaming is present in the running daemon — none were the cause.
• Transition broadcast is synchronous in Store::apply (no debounce), so a pushed state change reaches the stream in ms; the lag was entirely client-side sort staleness.

Test

• New unit test push_resorts_rows_so_a_state_change_moves_immediately.
• cargo fmt --all -- --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test -p muxa-cli all green (151 watch unit + 8 e2e).
• TUI not smoke-tested headlessly (needs a TTY); logic covered by the unit test + review.

🤖 Generated with Claude Code