Class RecognitionApi

The viewer-aware name for target as viewer would render it. See the class docstring for the algorithm. Always returns a string — the viewer-blind baseline is the universal fallback.

viewer is typed Stuff (not the narrower Sensor & Perception) so every call site — including the no-narrowing Mml render path — can pass whatever it has; the perceiver-ness is checked internally.

The single identity-learning write-sink. Records that viewer now knows subject as name (a non-null name from an introduction; a null name from a bare repeat-perception). Every recognition trigger funnels through here so they share one coalescing write.

No-ops when viewer can't hold beliefs or subject has no durable templatePath to key on.

The viewer-relative keyword set for targeting target — the tokens look <word> may resolve it by. Derived from the SAME perceived string describe renders, so the true name never leaks as a keyword. This is the Wave-5 name-leak gate.

Whether viewer already recognizes subject by a learned name. The gate the ambient auto-introduce triggers use so an NPC/player doesn't re-introduce to someone who already knows them. False when the viewer can't hold beliefs or has no name on record for the subject.

Generate a salient-feature description for a being the viewer doesn't recognize. Viewer-independent (features are objective; only the name is unknown). Reused by the viewer-relative targeting layer so naming and targeting can't diverge.