Like you, I also did a quick read to better understand Part 15, so let's compare notes.
Ignoring things like cordless phones, which have their own special carve-out allowing them high powered transmissions in the 902-928Mhz ISM band, I read the FCC rules governing the 902-928Mhz band as comprising 3 buckets rather than just the two you mentioned:
1. Spread spectrum and frequency hopping are allowed up to 1 watt. Governed by Part 15.247. However, RFM69x doesn't do spread spectrum, so it doesn't match that bucket. I'm not sure what potential, if any, it has for doing frequency hopping (though there is a section of the datasheet, Section 4.2.5, entitled " Optimized Frequency Hopping Sequences" which I haven't yet read). I need to look to into that more, but for now, until proven otherwise, it would appear the RFM69 doesn't fit this bucket.
2. Digital Transmission System (DTS). Also governed by Part 15.247. These devices are limited to a peak of 8dbm across any 3khz band. Also, the 6db bandwidth must be at least 500Khz. The Semtech paper I cited above is clearly trying to identify modes of operation (Tables 1 and 2) for the RFM69 to make the case that an RFM69 could qualify as a DTS if operated within those parameters. Why? Because the alternative is the third bucket (below), which appears to be generally more restrictive.
3. "Everything else" (as you call it). Governed by Part 15.249. The *average* Tx power is limited to no more than -1.12Dbm (roughly 0.77mw) and can have a *peak* of no more than 18.9dbm (about 78mw). Here's the rub: to calculate the average power using the rules of Part 15.249, you must determine what the worst possible duty cycle will ever be over a 100ms window. Applying a duty cycle correction factor based on that worst case duty cycle to the Peak Power yields a number which must be no more than -1.12dbm. So, the way I see it, you must either guarantee a very short worst-case duty cycle, or else your peak Tx power must be low, or some combination of both. Maybe there exist applications where the RFM69x can easily operate within those constraints, but if so, it's not intuitively obvious to me what they all might be. Speaking generally, the Part 15.249 rules don't appear to leave much wiggle room. That said, I suppose there's a chance you might get lucky when it comes to fitting your application within those constraints. For instance, maybe operating the RFM69 at a high bitrate while transmitting only a single very short packet (with, say, a 100ms rest interval between packets) might get you in or near the ballpark while still allowing a relatively high Tx power. Returning to my OP (above) and perhaps answering my own question, maybe the RFM69HW might outshine the RFM69W in that scenario and prove its worth all while maintaining compliance with the relevant FCC rules.