play::PlayTurnStore::open's current-schema arm and its final catch-all still
bail!ed unconditionally on a row it could not read, so a damaged or
future-schema row (any PlayTurn field-set change that lands without moving
STORE_SCHEMA) left the play card gone forever, exactly the PA.f177 gap but
for the arms PA.f177 itself did not touch. Both now retire through the same
retire_unreadable_row path the legacy arms use.
upgrade_legacy_row's own three non-decode exits (an inner schema tag that
disagrees with the envelope, and two compare-and-swap failure shapes) used to
be indistinguishable from a genuine decode failure once open caught the Err,
so a readable turn (Soul's probe: envelope v1, inner schema v2, decodes
cleanly) was destroyed under a sidecar reason that falsely claimed it "did
not decode". LegacyRowError now splits Decode from Other; only Decode
retires, Other propagates as an ordinary open failure that leaves the row
untouched.
Adds test coverage at the two construction sites PA.f179 fixed but never
pinned (open_question_of, and run's own StaleAnswer arm) proving a missing
question_token surfaces as None, not a fabricated Some(""); both fail under
Soul's exact mutation.
The retirement sidecar is now named *.msgpack instead of *.cc (PA.f192): it
is a one-shot forensic dump via std::fs::write, never a CultCache store, and
should not read as one. retire_unreadable_row's own doc comment now states
the write-before-destroy ordering and the orphan-sidecar risk explicitly
(PA.f189): a failed compare-and-swap after a successful sidecar write leaves
that sidecar in place with no corresponding successful retirement.
PlayTurnStore and PlayTable now remember the sidecar path a startup
retirement wrote (PA.f186's own plumbing; runtime.rs wiring in the next
commit).
Verified: cargo test -p ghostlight-dungeon --bin ghostlight-dungeon, 177
passed. Ran the mutations quoted in every new doc comment by hand and
reverted them; each failed the test it names (see the Cut 15 report for the
transcript).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>