After how many soft bounces, over what period, should a sender stop mailing an address?

Open question

RFC 3463 distinguishes permanent failures (class 5, "not likely to be resolved by resending") from persistent transient failures (class 4: mailbox full, greylisting, an overloaded receiver, a temporary policy block). A permanent failure suppresses the address at once; the transient class leaves the decision to the sender. Every mailing tool and provider applies some rule, such as n soft bounces within m days or k consecutive campaigns, but published reasoning behind the parameters is rare. Which rules have senders actually used, and with what outcome: how many suppressed addresses would have delivered later, how many recovered, how did bounce and complaint rates move afterwards? Does the right threshold differ between transactional mail, where a missed receipt matters to one person, and newsletters, where reputation matters to everyone on the list? Should a full mailbox (4.2.2) be counted like a temporary policy rejection (4.7.X), given that the first says something about the recipient and the second about the sender? And does the answer change when the same address bounces softly at one provider but delivers elsewhere, which points at a receiver-side condition rather than a dead mailbox?

What a useful answer contains

The sending volume and the stream type; the rule with its parameters and the status codes it counts; the observation window; the share of suppressed addresses that later delivered when retried; bounce and complaint rates before and after the rule was introduced; and whether the rule was chosen by evidence or inherited from a tool's default. Vendor defaults are useful when accompanied by their rationale.