fix(mercury): isolate reconnect generations

[samtsai]

🚀 Description, Motivation, and Context

Fixes a Mercury reconnect race that can leave a replacement WebSocket connected but unable to receive events.

Each Mercury connection now owns its listener, heartbeat loop, pong handler, and lifecycle channels. A stale connection cannot clear the connected state of a newer connection or trigger a second reconnect. WebSocket writes are also serialized across authentication, heartbeat, and shutdown paths.

This preserves the existing public API and automatic reconnect behavior while making reconnect state generation-safe.

📖 Issues and Related PRs

No issue or related PR was supplied.

🖼 Screenshots

Not applicable; this change does not modify the UI.

🌮 How This Was Tested

• Ran locally and verified
• Created tests (unit and/or E2E)
• Automated testing unnecessary (chores, doc updates, etc.)

Added a regression test that:

• Accepts an initial Mercury connection and confirms authorization.
• Closes the first connection to trigger automatic reconnect.
• Verifies the replacement connection sends heartbeat pings.
• Verifies the stale connection cannot mark the replacement disconnected.

Validation commands:

go test ./mercury
go test -race ./mercury
go test ./...
go vet ./...
git diff --check

Required Items Complete

• Unit tests written
• Integration tests written
• Coverage verified
• Functionally tested
• Deployable to production
• Uses feature flags
• README updated
• Build passes

Security

Not applicable. This change does not alter authentication, authorization, credentials, network endpoints, or stored data.