fix(tui): persist session before stall/cancel recovery so --continue keeps history (#2739) - #3285

Problem

Fixes #2739 (partial — the data-loss part) — after a long turn stalls and is recovered (or is cancelled with Esc), running --continue loads the previous session and the entire in-progress turn is lost.

Root cause

The stall watchdog and cancel paths clear turn bookkeeping without first writing the in-progress turn to disk. Three recovery paths in crates/tui/src/tui/ui.rs were affected:

• recover_stalled_runtime_turn (stall watchdog)
• mark_active_turn_cancelled_locally (Esc cancel)
• recover_engine_event_disconnect (engine disconnect)

Each finalizes the in-memory cells and resets turn state, but the partial turn lives only in app.api_messages and is never persisted. So --continue reads the last completed-turn save and the interrupted turn vanishes.

Fix

Add persist_recovery_snapshot() and call it on all three recovery paths before turn state is cleared. It uses the exact same mechanism as a normally-completed turn's auto-save:

if let Ok(manager) = SessionManager::default_location() {
    let session = build_session_snapshot(app, &manager);
    if app.current_session_id.is_none() {
        app.current_session_id = Some(session.metadata.id.clone());
    }
    persistence_actor::persist(PersistRequest::SessionSnapshot(session));
}

build_session_snapshot loads and updates the session keyed by current_session_id, so it writes back to the same session file --continue reads. This mirrors the existing completed-turn auto-save (see the SessionSnapshot persist at the top of the turn-complete handler), so it introduces no new persistence path or risk.

Testing

• New colocated tests issue_2739_stalled_turn_snapshot_preserves_api_messages and issue_2739_esc_cancel_preserves_session_messages_before_clear: verify the snapshot captures the in-progress api_messages and that cancel clears turn bookkeeping only after the messages are captured.
• cargo clippy -p codewhale-tui --all-features clean.
• cargo test -p codewhale-tui --all-features green (4857 passed / 0 failed).

Note

This addresses the session-loss symptom in #2739. The freeze/stall itself is already mitigated by the existing liveness watchdogs (reconcile_turn_liveness); this PR ensures those recoveries no longer cost the user their in-progress turn.