The shard's catalogue can now answer for 73 bodies it used to report absent —
they have no art at action 0 and real art at a later one, and their key says
which (`body/820/a23` is a horse). This side stores that action and stops
assuming `a0` anywhere.
The atlas join is the part that mattered. It read
a.asset_key = CONCAT('body/', b.body, '/a0')
which would have silently dropped exactly the creatures this phase adds. It now
reads the row's own action, with COALESCE for rows written before the column
existed — a NULL inside CONCAT makes the whole comparison NULL, which would have
taken every portrait off the site on upgrade with the database perfectly correct
and nothing in any log. It still matches at most one row per slug: a deeper key
(`body/820/a23/f4`) does not equal the catalogue key.
Verified against a real MariaDB with the live shard's own 1,095-row manifest: the
ALTER applies to an installed-shape table and is idempotent, the horse joins to
its a23 picture, a pre-phase-6 NULL-action row keeps its portrait, and a stored
frame key does not become a second candidate.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016wDDVXWMDz82WqE1i969r4
wtclaude
c6c51b190d