What idle timeouts do common NAT gateways and load balancers actually enforce, and what keep-alive interval survives them?

Open question

Long-lived connections (database pools, message-queue subscriptions, WebSocket sessions, VPN tunnels) cross NAT devices and load balancers that expire idle mappings. RFC 4787 sets a floor of two minutes for UDP mappings and RFC 5382 one of 2 hours 4 minutes for established TCP connections, but these are minimums for compliant NATs, not promises about the devices actually deployed; cloud NAT gateways, managed load balancers, home routers and mobile carriers are commonly reported to expire mappings much sooner, and the Linux keep-alive default of two hours (tcp_keepalive_time) is longer than any of those. When a mapping expires, neither endpoint learns anything: the next write is retransmitted into a void until the retransmission timer gives up, or a RST from the middlebox ends the connection abruptly, depending on the device.

The question is therefore twofold. Which idle timeouts are enforced today by widely used cloud NAT gateways, managed load balancers, home routers and mobile carriers, for TCP and for UDP separately? And which keep-alive intervals (per-socket TCP_KEEPIDLE settings, application-level pings, QUIC or WebSocket pings) have been shown to keep sessions alive across those devices without generating traffic that is itself a cost, for example on metered mobile links?

What a useful answer contains

• The device or service, its configuration where relevant, and the date observed, because timeouts change between product generations.
• The protocol and state observed: TCP established, TCP transitory (half-open or closing) and UDP each have their own timer in RFC 4787 and RFC 5382.
• The documented timeout, with the vendor page quoted and its URL, and the measured one if the two differ.
• Whether the device answers a packet on an expired mapping with a RST, an ICMP error or silence; silence is the case that keeps connection pools hanging.
• The keep-alive interval that worked and the shortest one that did not, and how the measurement was made, for example idle connections of increasing duration followed by a write, repeated enough times to rule out other causes.
• Anecdotes labelled as such and separated from measured results.