w6c+wwstage: reject untyped empty-[] alloc — require context, loud cannot-infer (#3 B', subsumes #5)

An empty `[]` carries no element type; ww gets it only from a let
annotation (the #45 retype). Both stages used to silently default the
element to u8, and in value-form positions (return / call-arg) the
lowering miscompiled — malloc(8) ignoring n, a 16B *u8|nomem where a 24B
slice was expected (#5). Now every empty alloc that isn't a
let-annotated binding fails to infer with a loud error, aligning ww DOWN
to harec (ref/harec/src/check.c:1801-1802).

Mechanism: clet / checkletassign flags the single alloc call node that a
`let x: []T =` rescues (save/restore around the init walk); the alloc
branch errors on any empty alloc that isn't that node. The #45 wide-T
retype path is kept. wwstage needs an extra not-yet-stamped guard because
resolvewalk re-types value nodes context-free after checkletassign.

Tests: negative cstage-driver 729 (table-driven: bare-let, return,
call-arg, assignment) + positive @test in attest_pass.ww exercising the
u8 and the wide-i32 (#45) paths at runtime. Both stages reject
symmetrically; byte-id verified on []u8 and []i32.

cmd/wcc/ww.h

@@ -547,6 +547,11 @@ struct Checker {
 			 * would shadow an imported module bareword. */
 	int	loops;	/* nesting count for break/continue */
 	int	errs;
+	Node	*alloc_octx;	/* #3/B': the one empty `alloc([], n)` call node
+				 * that has let-declared slice context this walk;
+				 * any OTHER empty alloc has no element-type hint
+				 * and must fail to infer (harec check.c:1801).
+				 * Set by clet around its cexpr, NULL elsewhere. */
 };
 
 void	check_init(Checker*, Arena*);