VACUUM INTO ':memory:' is refused; copy does not inherit pager sync flags

[timsehn]

Follow-up to #2373, which is closed: PR #2378 made in-transaction VACUUM error like stock and implemented file-backed VACUUM INTO. The seven vacuum-into testfixture gates that issue still justified are in place and still needed, so per check_testfixture_inventory_issues_alive.sh they need a live issue that describes the limitation as it stands now. Filed from #2400.

Remaining gaps after #2378

VACUUM INTO on a DoltLite-format source writes a compacted chunk-store copy to a file. Two SQLite behaviors are still missing:

1. VACUUM INTO ':memory:' (and expressions that resolve to it)

Stock attaches the destination, copies pages, then detaches. :memory: is a valid dest. DoltLite refuses it so the VFS does not create a literal file named :memory::

cannot VACUUM a doltlite database INTO an in-memory target

Verified 2026-08-24 on build/doltlite at d2c52f2953:

CREATE TABLE t(id INTEGER PRIMARY KEY, v INT);
INSERT INTO t VALUES(1,42);
VACUUM INTO ':memory:';
-- Error: cannot VACUUM a doltlite database INTO an in-memory target

The same error fires when the dest is an expression (VACUUM INTO (SELECT name FROM t2) with name = ':memory:'). File-backed VACUUM INTO 'copy.db' still works.

Gated as class=unsupported: vacuum-into-200, vacuum-into-300.

Making this work would mean copying the compacted image into a throwaway in-memory chunk store (stock discards the dest on DETACH anyway). The copy path today is file-only (gcWriteCompactedTo).

2. Destination does not inherit pager sync flags

Stock VACUUM INTO copies the source pager's synchronous / fullfsync flags onto the dest pager, so testvfs xSync counts match PRAGMA synchronous. DoltLite's copy has no pager to inherit from; dest writes go through the chunk-store file path. The dest therefore does not round-trip:

• PRAGMA synchronous = normal|full|off|extra (vacuum-into-710.2 .. 740.2)
• PRAGMA fullfsync = 1 with synchronous = full (vacuum-into-750.2)

Gated as class=intentional. Closing this would need dest writes to honor those PRAGMAs via VFS xSync the way the pager does — not a missing copy of user data.

Why the exceptions stand

Neither gap is a silent data-loss bug in file-backed VACUUM INTO. They are SQLite pager/memdb contracts the chunk-store copy does not have. The tests still fail on master; dropping the gates would redden vacuum-into.

Work this issue tracks

• Optional: copy into an in-memory chunk store so VACUUM INTO ':memory:' returns SQLITE_OK like stock, then drop the two unsupported gates and lower the ratchet.
• Leave the five sync-count assertions gated unless dest writes start inheriting pager sync flags.