fix(tui): add WebSocket heartbeat + auto-reconnect for silent drops (#32997)

[indigokarasu]

What / Why

Issue #32997: the Ink TUI (hermes --tui) WebSocket client had no heartbeat and
no auto-reconnect. A silent drop (macOS sleep, proxy idle timeout, VPN reconnect)
kills the TCP socket without a close event, so the client hangs forever —
input is blocked and only Ctrl+C recovers.

These four issues (#46563, #40164, #53142, #32997) share one root cause: the
WebSocket transport has no unified resilience contract. Three facets are already
covered by in-flight PRs and are intentionally NOT touched here:

• [Security][MEDIUM] No WebSocket message size limit — single client can OOM-crash the gateway #53142 server OOM → fix(tui-gateway): reject oversized websocket messages #53169 (message-size cap)
• Desktop chat can get stuck busy after idle WebSocket prompt timeout #40164 stale prompt timeout → [codex] Fix desktop stale WebSocket prompt timeout #40168 (shared-client stale-socket close)
• [Bug] Hermes Desktop: Session content disappears after WebSocket disconnect — pinned chat history lost #46563 Desktop session loss → fix(desktop): reattach sessions after websocket reconnect #45653 / fix(desktop): recover remote gateway reconnects #46040 / fix(desktop): preserve active chat on reconnect #53522 (Desktop reattach)

This PR closes the remaining facet: the TUI client now sends a periodic ping
keepalive, detects dead/silent connections, and auto-reconnects with exponential
backoff. Intentional kill() is honored (no reconnect loop).

Supersedes #30114 (exit-event-only restart, could not heal silent drops). Credit
to @sbiesaar for the original attempt.

How to test

• npm run typecheck --prefix ui-tui
• npm test --prefix ui-tui (gatewayClient connect/reconnect tests)
• npx eslint src/gatewayClient.ts src/__tests__/gatewayClient.test.ts

Platforms tested

Linux (Node undici WebSocket). Logic is transport-only; ping is guarded for
environments without WebSocket.ping.

[teknium1]

Thanks for targeting a real silent-drop gap in the TUI transport. Two correctness issues need resolution before this can be safely salvaged.

Problems

• ui-tui/src/gatewayClient.ts:242 treats absence of inbound application frames as a dead socket. tui_gateway/ws.py:319-337 sends gateway.ready then waits for client input, so a healthy idle client has no messages to refresh lastActivityAt. The local undici WebSocket declaration exposes no ping/pong API (node_modules/undici/types/websocket.d.ts:16-66), so the optional call at PR line 237 is not an acknowledged heartbeat. This will deliberately close healthy idle sockets; the new test at ui-tui/src/__tests__/gatewayClient.test.ts:497 currently asserts that behavior.
• ui-tui/src/gatewayClient.ts:350 schedules retry after the synchronous exit emission. The existing subscriber already invokes gw.start() for session recovery (ui-tui/src/app/useMainApp.ts:795-821), so the subsequently-created timer later replaces that fresh transport.

Suggested changes

• Add a server-supported, acknowledged heartbeat and test that healthy idle sockets stay connected.
• Use one recovery owner, and test an attached client with the current exit subscriber to prove a single reconnect occurs.

Automated hermes-sweeper review.

[indigokarasu]

Addressed the review in 383a000d6af33dee5b562e558b63014e7fa866d8.

1. Replaced the unacknowledged idle-timeout path with a JSON-RPC gateway.ping heartbeat. The client now waits for a server response before treating the socket as dead.
2. Added gateway.ping handling in the TUI WebSocket server.
3. Moved reconnect scheduling before synchronous exit emission. An immediate gw.start() subscriber now cancels the pending timer instead of creating a second transport.
4. Added regressions for healthy idle connections, missing heartbeat acknowledgements, and the exit-subscriber reconnect race.

Validation:

1. npm test --prefix ui-tui -- src/__tests__/gatewayClient.test.ts passed, 17 tests.
2. npm run typecheck --prefix ui-tui passed.
3. cd ui-tui && npx eslint src/gatewayClient.ts src/__tests__/gatewayClient.test.ts passed.

[indigokarasu]

@teknium1 this is ready for another look. I addressed the review feedback and moved the PR out of draft.

[100yenadmin]

The still-relevant TUI heartbeat and bounded reconnect behavior from this PR has been salvaged onto current main in #83166. Indigo Karasu's original commits are preserved as authored, with explicit credit in the combined PR. This cross-link does not close or rewrite this PR.