Offer advisory field-classification suggestions with mandatory human review

[jeremi]

Tracked by #590. Depends on #592 and should integrate with the review workflow in #596.

Outcome

Registry authoring tools can propose field concepts and classifications from bounded local metadata and synthetic fixtures, while keeping every proposal separate from reviewed Manifest metadata and active policy until a human explicitly accepts a generated patch.

Security and privacy rules

• Suggestions never grant access, activate policy, establish legal basis, infer consent, or decide eligibility.
• No raw production records, credentials, secrets, or environment contents are uploaded to a hosted service.
• Source inspection is local, bounded, explicit, and limited to the artifact classes selected by the operator.
• A suggestion remains a proposal even when confidence is high or a prior reviewer accepted a similar field.
• Missing classification is never silently converted into public or another permissive default.

Requirements

• Define a versioned proposal artifact containing the target field address, proposed concept or classification, immutable term IRI, basis, confidence or certainty class, tool/rule/model version, input artifact digests, and limitations.
• Support deterministic suggestions from field names, declared types, selected semantic profiles, schemas, and synthetic fixtures before considering model-assisted suggestions.
• Keep proposal generation separate from Manifest mutation. Acceptance produces an explicit reviewable patch.
• Require selected vocabulary/profile packages from Support versioned semantic profile packages and pinned vocabulary artifacts #591 and classification facets from Add field classification bindings to Manifest and Relay catalog metadata #592.
• Detect new or changed fields and quarantine them as unreviewed rather than inheriting an older field's classification by name similarity.
• Make rejection and correction useful feedback without turning prior decisions into automatic policy authority.
• Redact diagnostics and reports according to the existing project-authoring knowledge classification.
• Record whether a proposal was accepted, modified, or rejected without storing inspected source values.

Acceptance criteria

• A local synthetic workspace produces deterministic concept and classification proposals.
• No proposal appears in compiled Manifest or Relay policy until an explicit patch is accepted and revalidated.
• New fields remain visibly unclassified or unreviewed when no accepted binding exists.
• Proposal output includes basis, version, digest, uncertainty, and limitations.
• Tests prove raw fixture values, secrets, credentials, and environment contents cannot enter proposal output.
• The documented workflow covers accept, modify, reject, rerun, and stale-proposal behavior.
• Model-assisted operation, if added, is optional and cannot bypass the same artifact and review contract.

Non-goals

• Continuous scanning of production registries
• Hosted upload of person-level data
• Automatic policy activation
• Learning legal authority from prior approvals
• Inferring that an unknown field is safe because it has not been classified