`RetiredPlayTurnRow` wrote `rmp_serde` bytes to a `.msgpack` sidecar and
nothing in this Body could read it back — preservation in name only. It is
now JSON (the GameCult substrate doctrine's own xenos-boundary carve-out: an
operator opening a forensic dump by hand is exactly that boundary, not a
CultCache consumer), with the row's original bytes preserved verbatim as a
base64 field, `payload_base64`, and the sidecar extension renamed from
`.msgpack` to `.json` to match. A round-trip assertion on the existing
undecodable-bytes test decodes the sidecar and proves the recovered bytes are
byte-for-byte the original payload.
`PlayTable::retired_row_sidecar` cached the retirement fact once, at
`open`, keyed off this process's own store construction — so `playStatus`
on `/health` reported plain `ok` again after any restart, even though the
sidecar a prior process wrote was still sitting on disk with the only trace
of what happened. It is now a fresh directory scan on every read
(`retired_row_sidecars_on_disk`, filtered to this store's own filename
stem), so a restarted process still names the retirement until an operator
actually clears the sidecar file. A new test opens a second `PlayTable`
over the same store path after the first one's retirement and proves
`/health` still reports it.
Mutations run and reverted (not committed): `RetiredPlayTurnRow.payload_base64`
back to raw lossy text — killed by the round-trip test. `retired_row_sidecar`
back to a field cached at construction — killed by the restart test.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>