Even for FEC+interleaving without retransmission, the line profile sets a minimum INP value. Between the DSLAM and your modem, the actual INP level achieved may exceed the minimum. Without retransmission, it would likely be very costly in terms of the FEC overheads to wildly exceed the minimum INP value, so there probably tends to be little variance in the actual INP levels achieved. With retransmission, there's probably nothing much to lose by configuring the retransmission to provide the maximum amount of protection possible within the maximum delay and the available memory.
Bing. With that explanation, I find an answer to every query that I've had about the observed behaviour of reported INP values: that they are an "actual INP" rather than the "INP set by DLM" ... which is, as you mentioned, explicitly written in the specs as a "minimum INP" value.
Thanks @ejs.
It nicely dovetails with the observation that (with G.INP retransmission active), we do see minor variations in INP, even when a resync has been triggered at the customer end, not the DSLAM end.
Other observations are that the interleaving depth can vary between resyncs, and obviously depends on the sync-by-sync line conditions prevailing. In old-style FEC+interleaving, these values are high hundreds, low thousands. If adaption is necessary, they can change by a little. In new-style retransmission+FEC+interleaving, the choice appears to be either 4 or 8 - which is not very granular. It probably means that the modems are forced to add significantly more protection than is being strictly required.