Module census 

The load-time weights census: prove the checkpoint is the one we conformed against.

A wrong or stale checkpoint that loads “successfully” does not fail loudly — it produces plausible-looking garbage audio hours later, and the bug gets blamed on a kernel. So every load is checked against an expected (name, shape, dtype) inventory and any divergence is a named, itemized refusal at load time.

Three finding classes, all fatal by default:

• Finding::Missing — we need a tensor the file does not have. Certain failure downstream.
• Finding::ShapeMismatch / Finding::DtypeMismatch — the tensor exists but is not what we compiled kernels for. This is the class that silently produces garbage rather than crashing.
• Finding::Extra — the file has tensors we do not know about. Not dangerous on its own, but it is the signature of a different checkpoint (a sibling size, a newer revision), which is exactly what we want to catch before anyone spends a day debugging the audio.

The expected inventory itself comes from the OQ-2 tensor census over the pinned weights; this module is the mechanism, not the data.

Structs

CensusReport

The outcome of auditing one checkpoint.

ExpectedTensor

One tensor the manifest requires, as the OQ-2 inventory recorded it.

WeightsManifest

The expected tensor inventory for one checkpoint.

Enums

Finding

How a checkpoint diverged from the manifest.