An element generic declared on a reading input, absent from the output
and from every other input, has nowhere to go and nothing to be: the
node reads its declared attributes and never the element. That is an
attr-to-element map, so it accepts any upstream record wire.
The blessing is inferred rather than marked, since the signature already
says it. The two substrate pieces were both in place: the element rides
as the byte-carried token, so the registry takes one generic row instead
of a row per element type, and the reads resolve against the input's own
layout as they always have. Only the passthrough rule stood in the way,
in the three places that enforced it.
A generic the output carries, or one on an input with no reads, is
unchanged and still owes an implementations list.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The fold already holds the name and the read input's finished layout, so
the offset falls out there rather than at construction: `RecordLayout`
carries the resolved numbers and `set_layout` copies them into the read
slots. Constructors are untouched, and census-marker reads keep their
current installation.
A read meets the value type the name was written at, so a disagreement
between a read here and a write upstream is the same graph error as two
writes disagreeing; the one-name-one-type check now spans reads and
writes together. An absent attribute stays absent and the read serves
the forced default rather than reporting it.
`read_attribute` is the catalog's get half, typed and never `Option` at
the kernel boundary, with the name declared exactly as the write side
declares it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A name input declares where a placeholder's name is wired and nothing
more. The name is spent resolving the layout when the graph compiles, so
the kernel neither declares it nor is passed it; the return type's
`Named<X, V>` is the only tie between the write slot and the folded name.
The wire is untouched: the input keeps its declared position and crosses
as constant text, so a document's input order is unchanged. The
placeholder stays a concrete token, so it adds no generic for the node
to carry and none to go unconstrained.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
`Named<X>` in parameter position declares where a placeholder's name is
wired, and the macro gives that input constant text; `Attr<Named<X, V>>`
in the return writes under it. The kernel is handed the bare placeholder,
since the name is spent resolving the layout and a folded offset is all
the write needs, so the hot path matches a marker node's exactly.
A name-generic write names its value type through the wired generic,
which only an implementations row resolves, so those nodes emit their
layout meta per row rather than sharing one across rows.
`write_attribute` is the catalog's set half, restoring the identifier
master's documents carry with its input positions. Its name input is a
constant, so those documents resolve without migration, and the reset
marker that stood in for the missing node retires with it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
`compute_layouts` resolves every name-from-input write against the text
its name input carries, so a folded meta is indistinguishable from a
marker node's and the layout holds a `&'static str` from there on. That
leaves nowhere for a name to be computed, which is the whole of the
rule: a name input that is not a constant is refused right here, per
node, with the path the editor pins its diagnostic to.
One name carries one value type, checked across the census and the
names a node folds together, so no graph can declare one field at two
widths. `compute_layouts` returns those refusals rather than panicking.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
