Dock snapshot polling scales O(N) with every process the session ever observed #15

Risk (from the arc-4 delta review; not currently user-visible)

Since the #14 fix, the dock snapshot feed uses session.processes().list_all() on the ~250ms UI snapshot cadence (200ms deadline). At Lash 9debf5aab, list_all() → ProcessWorkObserver::observe_records awaits get_process_lease once per record, serialized on the same tokio_rusqlite connection live turns write through. Observer edges are deleted only on session deletion and lash-cli never calls prune_terminal_processes, so record count N grows monotonically for the session's lifetime (every subagent spawn adds one). On a deadline breach the snapshot returns processes: None and the dock silently stops updating.

Today N is small (a handful of subagent spawns per session), so this is a scaling cliff, not a live bug — but it is monotonic and silent.

Fix directions (either suffices)

• Drive terminal-row visibility off the ProcessChanged event stream instead of polling list_all() (poll list() for running rows only; terminal rows enter via events and expire via the existing 10s window), or
• Bound the poll: upstream ProcessListFilter gains a terminal-recency filter / bounded query (filed as Lash feedback), or CLI-side prune via prune_terminal_processes on a slow cadence.

Evidence

• crates/lash-cli/src/ui_effects.rs (collect_ui_snapshot → list_all()), snapshot cadence crates/lash-cli/src/interactive/mod.rs (~854), 200ms deadline crates/lash-cli/src/interactive/helpers.rs.
• Lash: crates/lash-core/src/process/observation.rs:234 (per-record lease await); crates/lash-sqlite-store/src/process_registry.rs:362 (edges deleted only with the session).