A database last opened by an old version can fail to open with 'database disk image is malformed'

[kevintseng]

What happens

A small database last opened by a version before 4.2.11 (or one holding Thai, Lao, Khmer or similar text indexed by an older build) fails every open after upgrading with database disk image is malformed. The database is not corrupt: an upgrade step removes old keyword-index rows using the new text segmentation before the index is rebuilt. Every surface (hooks, MCP, CLI, reindex --fts) fails the same way, so there is no in-product way out.

Steps

1. Take a small database written by an older build with CJK or Thai text and no title backfill yet.
2. Open it with 4.10.11.
3. memesh fails to start with database disk image is malformed, every time.

Expected

Upgrading never makes a readable database unopenable; the keyword index is rebuilt before any step depends on it.

[kevintseng]

Verified after merging #548 at a12c46a (tree 9afe90d11ab10d345f21a73346dff35a272c5644, identical to the reviewed candidate):

• A disposable legacy-index database containing CJK text opens through the built CLI; recall returns the original entity and observation.
• A second open succeeds and returns the same entity.
• The same scenario on the preceding main revision, 37fbb68, fails on both opens with database disk image is malformed.

This closes the upgrade/open-order defect. A separate existing limitation remains: recall against a read-only legacy index returns an empty result with degraded: false on both revisions. That limitation is not claimed fixed here and is being tracked separately. No live user database was repaired or rewritten for this verification.