Grade B: validate registry-sourced records before indexing

## StatGPT Backend version

latest

## What is the problem this feature will solve?

Part of #545. Extends #551, and builds on the shared validation gate introduced in #557.

Grade B records are derived from a registry's own metadata, so their failure modes differ from Grade C's. The registration guardrails in #551 bound a registry by its declared regional coverage and agency, but they act on the registry as a whole, not on each record: a registry that registers legitimately can still expose individual dataflows whose metadata claims coverage beyond that remit, or whose fields carry content unrelated to the dataset.

There is also no collection-side confirmation step here at all — nobody reviews a registry's dataflows before StatGPT reads them.

## What is the proposed feature or solution?

Validate Grade B records against the registry that supplied them, reusing the shared gate.

**Reuse the shared framework.** The validation gate, per-record isolation, and run reporting come from #557. This issue adds the checks specific to registry-sourced records.

**Per-record coverage and agency check.** Validate each dataflow's record against the registry's declared regional coverage and agency. A record claiming coverage beyond what the registry declared is rejected, so the guardrails hold per dataflow and not merely at registration.

**Malformed and misleading metadata.** Apply the same content checks as Grade C to the registry-sourced fields — the dataflow, its dataset-level attributes, and its DSD.

**Cost at scale.** A registry can hold thousands of dataflows, so per-record validation cost is a design constraint rather than a detail. Cheap deterministic checks should run first, with more expensive judgement applied only where they are inconclusive.

**Reporting.** Failures are reported per registry run alongside the per-dataflow read failures already isolated in #551, so one report covers everything that kept a dataflow out of the index.

**Not in this scope.** The registration guardrails themselves (#551), Grade B querying, and measuring how accurately validation classifies records.

## What alternatives have you considered?

- **Relying on the registration guardrails alone** — rejected; they constrain what a registry may register, not what each of its dataflows actually claims.
- **Reusing the Grade C checks from #557 unchanged** — insufficient; Grade C records are self-declared descriptions with no reference to check against, whereas a Grade B record can be checked against its registry's declared attributes. The jurisdiction check is stronger here and should use that.