`PlayTurnStore::open`'s current-schema arm routed every `Err` from
`read_current_schema_row` into `retire_unreadable_row`, but that function had
three exits and only one was a genuine decode failure: a canonical
re-encode mismatch or an inner `schema` tag disagreement both mean the row
decoded fine. Before this fix either case destroyed a perfectly readable
turn on the very next restart, the identical failure PA.f185 already fixed
for the legacy arms. `read_current_schema_row` now returns
`CurrentSchemaRowError::{Decode, Other}`, mirroring `LegacyRowError`: only
`Decode` reaches `retire_unreadable_row`, `Other` propagates as an ordinary
`open` failure and leaves the row untouched. Two new tests build a row that
decodes but fails each check and prove it survives open unretired.
Separately: `eve.rs`'s play card rendered "submit an empty message to
continue" hint for every `Running` turn, including the seconds a healthy
turn is genuinely running its own round loop. In that window `admit` fails
`run_lock`'s own `try_lock_owned` before inspecting anything and refuses —
contradicting the hint it had just shown. `PlayTurnView` now carries
`run_in_progress`, a non-blocking peek at `run_lock` taken alongside
`current_turn_view`'s other reads, and the hint gates on `state ==
Running && !run_in_progress`. `PlayError::TurnStillRunning` is split: the
lock-contention exit is now `TurnBusy` (its own accurate text), leaving
`TurnStillRunning`'s "only an empty continue is accepted" true for the one
case it still covers — text is non-empty on an otherwise-idle Running turn.
The `eve.rs:479`/`play.rs` doc comments both carried "PA.f188" for this
finding, colliding with the real PA.f188 (`open_question_of`'s construction-
site token coverage, `play.rs:9197`); both renumbered to PA.f193-D with a
note where the old number was.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
d082cd2277