perf(decode): parse struct destinations without the value tree

Assisted-by: GLM 5.3 Flash

docs/API.md

@@ -491,6 +491,23 @@ depth, including struct elements inside slices; map destinations accept every
 key by nature. When several keys are unknown, the message names the smallest
 one, so it does not depend on map iteration order.
 
+### Direct decoding
+
+For a struct destination whose type graph carries no untagged embedded map and
+no custom decode hook, `Unmarshal` and `(*Decoder).Decode` parse straight into
+the destination: the table skeleton is resolved against the struct schema while
+the document scans, and no intermediate value tree is kept. Values still flow
+through the ordinary assignment rules, so every conversion, hook and error the
+[Decoding](#decoding) section states holds verbatim; the parity with the tree
+path is pinned by a differential fuzz target that decodes every generated
+document both ways and compares the results.
+
+A document or destination the direct skeleton cannot model — an unknown table
+under strictness it must sink, a hook that needs the whole parsed value, an
+embedded map filler — falls back to the tree path and reruns, so the
+observable behaviour is always the tree path's, exactly. Nothing changes for
+`Parse`, `ParseMap` or the document API: the tree remains theirs.
+
 ### Cancellation
 
 `ParseContext`, `UnmarshalContext` and `(*Decoder).DecodeContext` accept a