fix(database): drop .failed state so clean health-monitor teardown stops alerting connection lost

[datlechin]

Summary

• ConnectionHealthMonitor.attemptReconnect no longer transitions to .failed after the while !Task.isCancelled loop exits. Loop exit means cancellation, which is clean teardown initiated by stopMonitoring(). The post-loop transition is replaced with a debug log and a plain return.
• HealthState.failed is removed entirely. It was unreachable through any non-cancellation code path.
• resetAfterManualReconnect() is removed: zero callers, and its only purpose was resetting from .failed.
• DatabaseManager+Health.swift drops the case .failed: arm that wrote session.status = .error("Connection lost") and called clearCachedData().
• The logLevel(for:) helper drops its dead .failed branch.

Net diff: +8 / -26.

Why this matters

This is bug R1 from the full-app audit. When the user closes the last window for a connection while the health monitor is in reconnect backoff, stopMonitoring() cancels the task. The reconnect loop exits cleanly, falls past the loop body, and unconditionally transitions to .failed. The session-level observer treats .failed as a real error, alerts the user that the connection was lost, and wipes cached schema/result data. The user's window-close action looks like a connection failure, and any concurrent session reuse can lose its cached data.

Removing .failed rather than guarding the transition is the correct cleanup: there is no path in the current code that produces a "real failure" distinct from cancellation. Adding a Task.isCancelled check before the transition would leave a dead state behind.

Test plan

• Open a connection, force a network outage so the monitor enters backoff, then close the window. Confirm no "Connection lost" alert and no cached-data clear.
• Cause a real ping failure followed by a successful reconnect handler call. Confirm the monitor still transitions through .reconnecting(attempt: N) and lands on .healthy.
• swiftlint --strict clean.

[chatgpt-codex-connector]

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