Turn training-data privacy and retention prose into a governed schema registry

[forkwright]

Finding

R8.4 requires versioned training event types with source, label, model, privacy class, and retention, but the concrete training taxonomy records only event, label, and model. Privacy and retention remain unbound prose rather than enforceable event metadata.

Evidence

• docs/requirements.md:82-92 requires the full versioned governance tuple.
• docs/training-data.md:7-17 defines the event table with only three columns.
• docs/training-data.md:19-35 restates privacy and retention principles but does not bind them to individual event versions or fields.

Why this matters

Emitters can add or change fields without a reviewable privacy classification, retention clock, deletion behavior, or compatible migration. Consumers cannot determine which version they received, and CI has no mechanism from which to derive completeness.

Required mechanism

Create a machine-readable registry containing, per stable event ID and version:

• field schema and source;
• training purpose/label semantics and allowed consumers;
• privacy class and redaction rule per field;
• retention trigger, duration, and deletion behavior;
• consent or other collection basis where applicable;
• compatibility and migration rules.

Require emitters to reference a registered version, fail CI for unregistered or incomplete event schemas, and test deletion/retention behavior from the registry.