The sidecar

A sidecar is a shared base plus one pattern-keyed variant. Two variants are KERNEL variants — vector and pure, the shapes a Bundle emits and the only shapes any loader launches. A third, neural_block, is a descriptor: it names kernels rather than being one, and nothing launches it (see The neural_block descriptor).

The base fields (compile.py::_base_sidecar) are present on every KERNEL variant; a descriptor carries a deliberately smaller base, spelled out in its own section below:

Key	Type	Meaning
format	"ptx" | "cubin" | "fatbin"	Artifact format the sidecar accompanies.
schema_version	int (1)	As for the manifest (backward-lenient / forward-strict).
pattern	"vector" | "pure" | "neural_block"	Selects the variant. Python’s LoadedVector/LoadedPure each refuse the wrong family outright (and default an absent value to "vector" for LoadedVector only). C++’s shared validate_sidecar treats it as a discriminant too — an unrecognized value is rejected, naming the supported set — but never dispatches a variant class (every C++ binding decision is driven by arg_spec alone), so it refuses only an unrecognized value, not a “wrong-for-this-caller” one, and stays absence-lenient.
aether_abi	const "aether-abi/1"	Binary-ABI tag (see The aether_abi tag is separate).
scalar_type	"float64" | "float32" | "softdouble"	Real scalar type the kernel was compiled for.
kernel	string	The extern "C" entry-point symbol to resolve in the artifact. Always "raptor_kernel" (eagle.KERNEL_NAME): raptor’s schema fixes this one name for every producer, hand-written kernels included, so the artifact file name, not the symbol, tells kernels apart.
vector_inputs	array[string]	Named vector inputs, in binding order.
params	array[string]	Named uniform scalar parameters.
per_sample	array[string]	Named per-sample scalar inputs.
arg_spec	array[[role, name]]	The ordered [role, name] binding list — role first (see The arg_spec role vocabulary). Present on every variant.

Variant additions

vector adds:

Key	Type	Meaning
accumulate	bool	Whether the out sink uses the outVec += accumulate idiom.
sink	string	The accumulate-sink contract name (e.g. "outVec-scratch") — a per-kernel deterministic-reduce scratch slot, not the final total.
vec_widths	object{string→int}	Per-name vector width (component count) for host (W, N) allocation.
buffers	array	Declared read-only lookup buffers (may be empty; see below).

pure adds:

Key	Type	Meaning
mutables	array[object]	The writable per-sample state that is the pure output. Each element is {name, dtype ("float"|"int"|"vector"|"matrix"), width, default (number|null)}, plus shape: [R, C] only when dtype == "matrix".
vec_widths	object{string→int}	Per-name vector width (includes vector Mutables).
buffers	array	Declared read-only lookup buffers (may be empty).
matrix_inputs	array[string]	Only when the kernel has matrix inputs: the mat_in binding order.
mat_shapes	object{string→[R, C]}	Only when the kernel has matrix inputs: the (R, C) shape per matrix-input name.

Each buffers element (compile.py::_buffers_field) is {name: string, kind: "lookup", dtype: "float", shape: array[int], count: int}, where count == prod(shape) is the flat element count the registry uploads once at consolidation and binds by value as a flat handle.