fix(unlockmode): Exit Without Saving now reverts a moved fallback anchor

[dfrisone]

Reported by @grouch on 8.7.6: throw a fallback anchor far away to reach something underneath it, then Exit Without Saving, and it stays where it was thrown. This is ours, not an interaction with another addon.

Cause

RevertPositions restored nine anchor fields from the snapshot and then deliberately re-attached the LIVE fallback on top.

That carry-over existed for a good reason: SnapshotPositions never captured fallback, so without it a revert would have destroyed every fallback, including ones set in earlier sessions. The side effect was that a fallback edited during the session also survived the discard, which is exactly the report. The anchor reverts, and then the edited fallback is put back over it.

Fix

Snapshot fallback by copy, restore it from the snapshot by copy, and drop the carry-over it made redundant.

The copy is load-bearing. Fallbacks are positional, and _NudgeSelectedFallbackGhost does

fb.offsetX = (fb.offsetX or 0) + dx
fb.offsetY = (fb.offsetY or 0) + dy

so it mutates the table in place. A snapshot holding a reference would be dragged along by the edit and the revert would be a silent no-op. The restore copies for the same reason in the other direction, so the next session's nudges cannot reach back into the snapshot.

Pre-session fallbacks are still preserved, which is all the carry-over was protecting. Session edits now revert like every other position.

Why dropping the carry-over is safe

Both the old carry-over and the new snapshot iterate pairs(anchorDB), so coverage is identical rather than merely similar. Entries created during the session are dropped by both, since an entry with no snapshot is not recreated at all, which preserves the existing rule that a link reverting away takes its fallback with it. Pre-session fallbacks survive in both.

fallback only ever holds four scalars (target, side, offsetX, offsetY), with no frame references and no nesting, so CopyTable has nothing deep or cyclic to walk.

One deliberate widening

Removing a fallback mid-session now reverts too. Previously that survived Exit Without Saving, because the carry-over simply had nothing to re-attach. It is now discarded like any other session edit, which is what Exit Without Saving is supposed to mean, but it is behaviour beyond the reported case and worth knowing about.

Performance

SnapshotPositions has one caller, at unlock-mode entry. Not per frame, no timers, nothing added to any update loop. The change costs one four-field table copy per anchored element, once, when unlock mode opens. The same function already copies the whole unlockAnchors, unlockWidthMatch, unlockHeightMatch, and phantomBounds tables a few lines later, so this is a rounding error against work already being done, and CopyTable is the established idiom in this file.

Ghost visuals need no extra handling: RevertPositions already ends by hiding the fallback ghosts, and they recompute from fb.offsetX/offsetY whenever they are next shown.

Notes

One file, net 16 insertions and 20 deletions, so it removes more than it adds. No new strings, no options changes, no saved-variable format changes.

Tested in game: a fallback anchor thrown during a session now returns to where it was on Exit Without Saving.