The hint as the model is shown it, with the engine's own accessor.
The key NAMES are gated. A map with a stable key set - per-user scores, per-tenant flags - has perfect
recurrence and so scores maximally as a record; {"ZZALICE":3,"ZZBOB":5} on every row is structurally
identical to {theme,notify} on every row. No threshold separates them, so the names ride the same
opt-in as every other cell value and the default states only how many there are. The accessor is
always safe and is what stopped the model reaching for LIKE, which was the larger win.
The key list is trimmed at a key boundary: the comment cap appends an ellipsis, which would cut the
last name in half and offer a key that does not exist.
The hint as the model is shown it, with the engine's own accessor.
The key NAMES are gated. A map with a stable key set - per-user scores, per-tenant flags - has perfect recurrence and so scores maximally as a record;
{"ZZALICE":3,"ZZBOB":5}on every row is structurally identical to{theme,notify}on every row. No threshold separates them, so the names ride the same opt-in as every other cell value and the default states only how many there are. The accessor is always safe and is what stopped the model reaching for LIKE, which was the larger win.The key list is trimmed at a key boundary: the comment cap appends an ellipsis, which would cut the last name in half and offer a key that does not exist.