Module bounds 

ONNX spec-pinned constants.

ADR-060 removed the fixed-width two-level commitment (ONNX_CANON_MAX_BYTES / ONNX_CANON_BYTES) and the application-policy capacity profile (OnnxHostBounds / OnnxAddrBounds) with its node-count / initializer-count / attribute-count / IO-count / tensor-data ceilings. The realization now emits the full flat canonical skeleton (the ModelProto structure emitted inline, with variable-length leaves — tensor data, strings, opaque sub-message payloads — replaced by their SHA-256 digests) as an unbounded alloc buffer that flows through the pipeline as a borrowed carrier. Every count and width is unbounded.

What remains are ONNX spec / policy constants (the admitted IR version range, the default-domain opset-version minimum) plus one native-stack-overflow guard on the recursive subgraph descent.

Constants

ONNX_IR_VERSION_MAX

The highest ONNX IR version this realization admits — onnx.proto’s current Version::IR_VERSION (= 13, 2026). The realization accepts any ir_version in 1..=ONNX_IR_VERSION_MAX: the canonical skeleton is IR-version-agnostic (the field numbers it reads are stable across IR revisions; IR-v10+ NodeProto.overload simply reads empty on older models), and the ir_version value itself is bound into the skeleton, so two IR revisions of the same logical model canonicalize distinctly. Admitting the range lets the realization content-address real-world models (published exports are predominantly IR 6–10) rather than only the latest revision.

ONNX_OPSET_VERSION_MIN

Policy: the minimum opset version accepted for the default domain "". ONNX mandates no minimum (= 1 accepts any opset); raise per application policy. Inlined from the pre-ADR-060 OnnxAddrBounds profile (ONNX_OPSET_VERSION_MIN = 1).

ONNX_SUBGRAPH_DEPTH_MAX

Native-stack-overflow guard on the recursive subgraph descent (If / Loop / Scan bodies via GRAPH / GRAPHS attributes). Guards the call stack against pathologically-nested subgraphs; it is not a ceiling on node / attribute count at any level.