Lorin sent me detailed explanation the other day, on request, about how the logic for versions work (reposted here as it’s so useful) 
What makes Serento Call two albums “versions”.
The short answer to “what drives this” is that there is no single rule. There are several independent passes, they run one after another, and they only ever MERGE. Two albums end up in one version group if ANY pass links them, and nothing that a later pass sees can un-link them. That is deliberate: a duplicate tile on screen is annoying but recoverable, while a wrongly merged album gets hidden behind its group winner, and a hidden album is not recoverable by the user. So every rule below is written to err towards leaving things apart.
Here are the passes that matter for your cases.
Pass 1: The folder on disk
I find the album folder by POSITION, not by searching the path. Starting at the file, I walk upwards past any component that reads as a disc subfolder, and the first component above those is the album folder. A “disc subfolder” is one of cd, disc, disk, sacd, dsd, side, part, vol, volume followed by a number, with an optional space, underscore or hyphen in between. So CD1, CD 1, Disc 2, Disk-3, SACD 1, DSD 2, Vol 3, Part 1 and Side 1 all qualify.
Then: if the album folder’s own name contains a dash separator (" - ", en dash or em dash), the key is that folder name on its own. If it does not contain a dash, the key becomes parent folder + “/” + album folder, so that two unrelated albums each sitting in a folder called “Live” or “Remaster” cannot collide with each other.
Pass 1 ignores the artist tag completely, on purpose. In classical libraries one server tags by composer and another by performer for the very same files, so on this pass the path is the truth and the metadata is an opinion.
That is your observation number 1. Putting an album in a subfolder only makes it “a version” of the album above it when that subfolder’s name matches the disc pattern above. A subfolder named “5.1” or “Hi-Res” or “Japanese” does not match, and it gets its own separate key.
Pass 2: Artist + Title, after stripping trailing qualifier blocks
I bucket by artist plus a normalised title, where the normalised title has trailing parenthesised or bracketed blocks removed. Whether a block is removed comes down to one rule, and this is the rule that explains your inconsistency:
- A block with NO digits at all is always stripped. So [DSD], [Remaster], (Deluxe Edition), [Stereo], (Mono) all disappear and the albums group.
- A block with digits AND letters is stripped only if one of its words is a recognised qualifier. The vocabulary is: remaster, remastered, remasters, remastering, reissue, reissued, edition, editions, anniversary, deluxe, expanded, mono, stereo, remix, remixed, version, dsd, dsf, sacd, mqa, hdcd, xrcd, hires, hi, res, bit, bits, khz, hz, kbps, cd, lp, flac, wav, aiff, alac, vinyl, rip.
- A block with digits and NO letters is KEPT, unless it looks like a sample-rate pair containing a slash such as (96/24).
So “[5.1]” is digits with no letters and no slash, which means I keep it, which means “Blood on the Tracks” and “Blood on the Tracks [5.1]” get two different title keys and this pass never links them. The reasoning behind that rule is that a bare number in a trailing block is far more often a year, a volume or a disc number than a format marker, and two of those really are two different releases. 5.1 lands in that bucket by accident, and I agree that is unfortunate.
There is one more door for square brackets specifically: a block is also stripped when it reads as a record label catalogue code, meaning it contains at least one all-uppercase token of two or more letters and at least one token carrying digits, with no four-digit year and no part word. That is why “[MFSL UDCD 2-537]” is stripped and groups correctly. “[5.1]” has no uppercase code, so that door does not open for it either.
The practical consequence for you, and it is a small tagging change rather than a restructure: put a recognised word inside the block. “[5.1 Edition]” groups. “[Multichannel Edition]” groups. “[SACD]”, “[DSD]”, “[Stereo]”, “[Mono]”, “[Remaster]”, “[Deluxe Edition]” all group. “[5.1]” on its own does not.
Pass 3: The fingerprint
If neither of the above fires, I still try same artist + same track count + one title being a clean prefix of the other + total running time within a tolerance. This is why some of your pairs group with no qualifier word anywhere: a 5.1 transfer normally has exactly the same track list and near-identical runtimes, so the fingerprint catches it even though the title block did not.
And this is exactly why the behaviour looks inconsistent to you. Take The Division Bell: the original and the first remaster have the same track count and nearly the same runtime, so the fingerprint links them. The anniversary edition has extra material, so its track count differs, the fingerprint cannot fire, and the title block is not a recognised qualifier either, so nothing links it and it stands alone. The same reasoning applies to your four copies of Wish You Were Here: the two that appear together are the two whose track lists match. The Endless River pair is the [5.1] case from pass 2.