perf(terminal): coalesce PTY output posts and unblock the keystroke path

[AThraen]

Fixes the typing-freeze reported from live use: with many live sessions, typing into a pane could lag by minutes before the characters appeared.

Closes #70.

Root cause

Three compounding causes, all on the shared WPF UI dispatcher. Full trace in #70; summary:

1. One dispatcher post per PTY chunk. TerminalBridge.OnPtyData did a Dispatcher.BeginInvoke for every chunk from every session. Chatty background sessions (Claude spinners, status-line repaints) flooded the queue.
2. Input was starved too, not just rendering. CoreWebView2.WebMessageReceived is raised on the UI thread, so a keystroke waits in the same backlog it's competing with — then the echo waits again to render. Two full queue traversals per visible character, which is how hundreds of ms becomes minutes.
3. O(N) work per keystroke. AlertCount (Sessions.Count(s => s.NeedsAttention)) was recomputed twice per keystroke with WPF binding invalidation each time.

Scale context: this was originally measured at 23 sessions; the reporter's state.json has 39.

Changes

• New OutputCoalescer (Terminal/OutputCoalescer.cs) — chunks arriving before the dispatcher gets a turn collapse into a single post. Scheduler and emitter are injected, so the state machine is unit-testable with no dispatcher and no WebView2.
• Active-session priority — new TerminalBridge.IsForeground, driven from UpdateActiveTerminalHighlight. Background sessions post at DispatcherPriority.Background; only the active pane posts at Normal.
• Keystroke path — removed the redundant AlertCount raise in the UserInput handler (AlertCleared fires unconditionally from NotifyUserInteracted and already raises it). The remaining raise goes through RaiseAlertCountIfChanged, which suppresses no-op raises. AlertRaised/AlertCleared move from blocking Invoke to BeginInvoke.
• The pre-ready output buffer now drains through the coalescer as well, so load-time output and post-ready chunks share one buffer and keep arrival order.

MainViewModel's GroupsChanged Invoke is deliberately left blocking — low-frequency group CRUD, not on the hot path, and deferring it could reorder sidebar rebuilds.

Verification

• 7 new unit tests for the coalescer (single/many appends, re-entrancy during flush, concurrent appends losing no data), written test-first and watched fail
• 213/213 unit tests pass
• App builds and starts clean

Not yet verified: the latency improvement itself is unmeasured under real load. That needs a multi-session run with Settings > Diagnostics > Trace terminal input/output on, comparing OUTPUT dispatcher latency before/after. Worth doing before this is assumed fixed.

Follow-ups (deliberately not in this PR)

• Git polling — every session polls git status --porcelain every 10s forever regardless of visibility (SessionViewModel.cs:109). At 39 sessions that's ~8 process spawns/sec of full working-tree scans. Steps 4-5 in perf: coalesce per-session PTY output dispatcher posts to prevent UI starvation #70.
• AlertDetector.Feed reallocates a System.Threading.Timer per output chunk.
• perf: session restore on startup is fully sequential (~131s for 25 sessions; launch cost, not sleeps) #82 — startup/shutdown latency, same scale-with-session-count theme.

🤖 Generated with Claude Code