Autonegotiation is, as you suspect, capability negotiation and not training / based on the signal integrity conditions whatsoever. In fact, since autoneg occurs over the two 'normal' pairs (used in 10base-T and 100base-TX), it's possible to successfully negotiate 1000base-T even if the other two pairs are completely missing, and the link will be completely nonfunctional. Some PHYs do include cable tests / TDR measurements, but this isn't generally done routinely before link-up, and isn't standardized or part of the autonegotiation process.
In essence, it’s two 500Mbps links aggregated. So chances are that degraded half-gigabit times two is better than dropping to a single 100Mbps, at least if the links half work.
Not quite, all 4 pairs are used bidirectionally in 1000base-T; it's actually 250Mbps x 4 (same 125Mbaud as 100base-TX). Also the line coding splits each byte across all 4 channels, so if you lose any of them, the link won't pass any data at all. Even if it did, higher level network protocols would tolerate 50% packet loss extremely poorly, and real-world performance would be much, much, much worse than a properly working 100mbps link (it would be essentially unusable).
Due to improvements in line coding, the use of bidirectional channels, and adding more pairs, the signal quality requirements for 100base-TX and 1000base-T are basically the same, so as long as the extra pairs aren't damaged, it shouldn't generally have a different error rate than 100base-TX on the same cable.