v3.20.2 — unapply accepted a squash it ignored

Fixed. 3.20.1 removed squash from unapply's call site and its docstring but not from its signature, so the published wheel takes unapply(values, squash="hard") and quietly does the soft inverse. Nothing downstream is wrong — the two inverses agree everywhere a hard clip can produce a value — but a parameter that is accepted and ignored is the exact defect 3.20.0 and 3.20.1 are about, and it shipped inside the release that was about it.

It survived because the edit removing it used a string replacement with no assertion on the match, and the signature's line wrapping had changed under an earlier edit in the same pass. Tests passed, ruff passed, the source read correctly everywhere except one line. What caught it was checking the published wheel rather than the working tree.

test_unapply_takes_no_squash_and_does_not_pretend_to pins it as an interface fact, and the 3.20.1 changelog entry carries the correction rather than reading as though the cut had landed.

Upgrade from 3.20.1 if you call unapply.