perf(tui): off-load snapshot refresh from event loop

[Roberdan]

Problem

cvg dash is unusable on real workloads. With 50 plans / 351 tasks the
dashboard shows a 1-2s black screen at startup, freezes for ~1s every 5s
during the periodic refresh, and lags up to 200ms per keystroke.

Root causes (crates/convergio-tui/src/lib.rs:120 and friends):

• state.refresh(&client).await is called before the first
  term.draw, so the user sees nothing until the first snapshot — and
  the snapshot is dominated by gh pr list --state all --limit 100
  (~1s on a warm cache).
• The same refresh().await sits inside the tokio::select! event
  loop, so every tick / r keystroke blocks key handling and rendering.
• gh pr list ran on std::process::Command::output(), pinning a
  runtime worker for the entire shell-out.
• EnableMouseCapture was on with no mouse handler — it stole native
  scroll and spammed the input poll with Noop events.
• event::poll(200ms) introduced up to 200ms of input latency.

Why

Dashboard responsiveness regression is felt on the first session of the
day, the loop tick, and every keypress. Constitution P3 asks the TUI
to be usable on a basic 80×24 terminal — it must also be usable in real
time.

What changed

1. Decouple Client::snapshot() from the render loop via mpsc.
   First paint shows a skeleton frame immediately; data folds in as it
   arrives. Periodic tick and r spawn a tokio task; an in-flight
   flag debounces overlapping refreshes. Adds AppState::apply_snapshot
   so the loop can push a pre-fetched Snapshot without re-awaiting.
   The new method (and the existing refresh wrapper) live in a new
   state_lifecycle module to keep state.rs under the 300-line cap.

2. gh pr list runs on tokio::process::Command and joins the HTTP
   fan-out via tokio::join!. Results are memoised for 30s — most
   refreshes pay zero gh cost. --limit reduced from 100 to 50.

3. Drop EnableMouseCapture, shorten the key-poll window from
   200ms to 50ms.

Validation

Measured on the live dashboard (50 plans, 351 tasks, daemon on loopback):

State	Before	After
First paint	~1.5s	<100ms
Refresh freeze	~1s	0 (async)
Cached refresh	~1s	~20ms
Keystroke lag	≤200ms	≤50ms

Snapshot bench against the live daemon (release):

Run 1 (cold gh + cache miss):     1.6s   prs=50
Runs 2..6 (cache warm):           ~20ms  prs=50

Local pipeline:

• cargo fmt --all -- --check — clean
• RUSTFLAGS="-Dwarnings" cargo clippy --workspace --all-targets -- -D warnings — clean
• RUSTFLAGS="-Dwarnings" cargo test --workspace — 1043 passed, 0 failed
• pre-commit hooks: file-size, fmt, clippy, context-budget all green

Impact

• convergio-tui only — read-only client, no daemon-side change.
• New public method AppState::apply_snapshot(Result<Snapshot>).
  AppState::refresh(&Client) is preserved for the existing test path.
• PR list scope shrunk from 100 to 50; the dashboard already truncates
  the rendered list well below 50.
• Tradeoff: PR data is now up to 30s stale (cached). r still triggers
  a fresh refresh; PR pane labels show last refresh timestamp via the
  existing footer.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.