feat(schemas): measurement capability block + brand.json metric_categories (#3612)

[bokelley]

Summary

Adds a measurement capability block to get_adcp_capabilities so measurement agents self-describe their per-metric catalog — the same pattern every other AdCP agent type already uses (sales/creative/governance/brand/buying/signals/rights). Adds optional metric_categories[] to brand.json's brand_agent_entry, paralleling rights agents' available_uses[] / right_types[] for coarse pre-call filtering by AAO.

Closes #3612 (the protocol surface piece of #3586's per-metric catalog discovery).
Unblocks the buyer-proposed vendor-metric flow in #3576 — buyers can now know which vendor metrics to propose on package-request.committed_metrics. Without this PR, the request-side vendor-scoped entries land partially blind.
Unblocks #3613 (AAO crawler + index implementation) — schema for the data the crawler ingests.

The architectural anchor

Every AdCP agent type publishes capabilities at the agent itself via get_adcp_capabilities. brand.json's agents[] array is just the directory; capability info lives at the agent. The one partial exception is rights agents, which put available_uses[] and right_types[] on brand.json for coarse filtering — full pricing/terms still come from the live agent.

This PR follows the rights-agent pattern for measurement — and per protocol-expert feedback, models the capability as a block (presence = support) rather than a supported_protocols value. Measurement agents have one surface (this catalog), not a tool-set with mandatory tasks the way media_buy / signals / governance do. Same precedent as compliance_testing / webhook_signing blocks.

Layer	Where	Purpose
Coarse filter	brand.json agents[type='measurement'].metric_categories[]	Cheap directory query — "any attention vendor?"
Canonical catalog	Agent's get_adcp_capabilities.measurement.metrics[]	Full per-metric data with id, category, standard, accreditations, unit, methodology
Cross-vendor index	AAO /api/registry/measurement-vendors	Crawled aggregation of capability responses on a TTL (separate PR, #3613)

Schemas added

• enums/measurement-category.json — closed 12-value enum (attention, viewability, invalid_traffic, brand_safety, brand_lift, incrementality, audience, reach, creative_quality, emissions, outcomes, other). Includes viewability/invalid_traffic/brand_safety (MRC/TAG/GARM) per expert review.
• get-adcp-capabilities-response.json — new measurement block with metrics[]. Each metric carries metric_id and category (required), plus optional standard_reference, accreditations[] (third-party certification — distinct from standard_reference), unit, description, methodology_url, methodology_version. additionalProperties: false with explicit ext slot. uniqueItems: true on the array.
• brand.json brand_agent_entry — optional metric_categories[] array.

Why accreditations[] is separate from standard_reference

A metric can implement a published standard (URL pointing at the spec) without holding independent third-party accreditation. Buyers asking "is this MRC-accredited?" need a structured answer that survives URL parsing — every vendor pasting the same MRC URL whether accredited or not gives a false signal of comparability. The split surfaces the distinction at the schema layer.

Doc updates

• docs/protocol/get_adcp_capabilities.mdx — new measurement section with field table, response example showing accreditations[] and methodology_version, the discovery-vs-settlement framing, and an explicit "this is a discovery surface, not a rate card" callout (pricing/SLAs/coverage are negotiated per buy via measurement_terms).
• docs/registry/index.mdx — refined measurement-vendor discovery section that points at the new capability block and forward-references the AAO index endpoint (AAO crawler + index for measurement-agent capabilities #3613) and the buyer-agent direct-call docs (Buyer-agent direct-call vs AAO-index pattern for measurement capabilities #3614).

Worked example (locked into the test suite)

{
  "supported_protocols": ["sponsored_intelligence"],
  "measurement": {
    "metrics": [
      {
        "metric_id": "attention_units",
        "category": "attention",
        "standard_reference": "https://iabtechlab.com/standards/attention-measurement",
        "accreditations": [
          {
            "accrediting_body": "MRC",
            "certification_id": "MRC-ATT-2026-001",
            "valid_until": "2027-12-31",
            "evidence_url": "https://mediaratingcouncil.org/accreditations/attentionvendor"
          }
        ],
        "unit": "score",
        "description": "Eye-tracking-based attention score (0-100). Computed from a panel of opted-in households.",
        "methodology_url": "https://attentionvendor.example/docs/attention-units",
        "methodology_version": "v2.1"
      }
    ]
  }
}

Backwards compatibility

All additions are optional and additive. Sellers without measurement capability are unchanged. Measurement vendors gain a structured catalog surface.

WG review

Hybrid design reached through the discussion thread on #3586 → #3612. Three independent expert reviews shaped the final shape: moved measurement out of supported_protocols (capability-block pattern), added missing categories, added methodology_version, added structured accreditations[], switched to additionalProperties: false with explicit ext, locked an example into the test suite, added uniqueItems: true on metrics[]. Labeled needs-wg-review.

Test plan

• npm run build:schemas — clean
• npm run test:schemas — 7/7
• npm run test:examples — 36/36 (added 2: positive measurement example + negative duplicate-rejection)
• npm run typecheck — clean

Closes #3612.

🤖 Generated with Claude Code