feat(catalog): project publisher_properties fan-out into catalog_agent_authorizations (#4841)

[bokelley]

Summary

Closes the catalog snapshot gap from #4840's rollout. Adds a new adagents_authoritative evidence value to catalog_agent_authorizations, lifts the cross-publisher refusal specifically for the fan-out shape, and backfills the ~6,800 existing cafemedia rows. /registry/authorizations and /registry/authorizations/snapshot (partner-sync endpoints, catalog-only readers) now see the manager-asserted children.

Why the cross-publisher guard stays

publisher-db.ts:968 refuses to project catalog_agent_authorizations rows where the selector's publisher_domain ≠ the publisher being projected. This is load-bearing for catalog trust — without it, a hostile manager file could project authoritative rows for publishers that didn't authorize anything. Don't lift the guard.

Instead, give the fan-out its own evidence value:

Evidence	Source	Trust
adagents_json	publisher's own /.well-known/adagents.json	strongest — bilateral
agent_claim	sales-agent list_authorized_properties	lower — unverified peer
community	AAO moderator curation	curated
adagents_authoritative (new)	manager file (e.g. cafemedia) names this publisher in publisher_properties[].publisher_domains[]	medium — manager-asserted, no bilateral

Consumers that want bilateral-only filter evidence=adagents_json; consumers that want the full picture take both.

Files

• Migration 488 — CHECK widening + backfill from publishers WHERE discovery_method='adagents_authoritative' ⋈ agent_publisher_authorizations WHERE source='adagents_json'. ON CONFLICT DO NOTHING.
• publisher-db.ts — new recordCatalogFanoutAuthorization({ agentUrl, childDomain, authorizedFor? }). INSERT-or-UPDATE with the new evidence value, created_by='system', no property_rid.
• crawler.ts — fan-out helper calls the new writer per (agent, child) pair alongside recordChildPublisherFromManager (publishers row) and recordAgentFromAdagentsJson (legacy authz row).
• federated-index-db.ts — 6 evidence-CASE statements add WHEN 'adagents_authoritative' THEN 'adagents_json' so the dual-read source field round-trips correctly.
• Integration test (catalog-fanout-projection.test.ts) — 6 cases: row shape, idempotent re-write, coexistence with adagents_json for same (agent, publisher), invalid agent_url silent-skip, canonicalization, view round-trip.

Production verification after deploy

# Re-trigger cafemedia crawl (deploy runs the backfill as release_command).
ADMIN_KEY=$(grep "^ADMIN_API_KEY=" .env.local | cut -d= -f2)
curl -X POST 'https://agenticadvertising.org/api/registry/crawl-request' \
  -H "Authorization: Bearer $ADMIN_KEY" -H 'Content-Type: application/json' \
  -d '{"domain":"cafemedia.com"}'

# Wait ~60s, then probe the catalog-only auth snapshot endpoint:
curl -sS 'https://agenticadvertising.org/api/registry/authorizations?agent_url=https://interchange.io&evidence=adagents_authoritative' \
  -H "Authorization: Bearer $ADMIN_KEY" | jq '{count: (.rows | length), evidences: ([.rows[].evidence] | unique)}'
# Expected: ~6,800 rows, evidence=["adagents_authoritative"]

Test plan

• Typecheck passes
• 6 integration tests for the new writer
• Reviewer: backfill query plan on production catalog (~thousands of rows expected)
• Reviewer (post-deploy): partner snapshot endpoint returns ~6,800 cafemedia children

🤖 Generated with Claude Code

[aao-release-bot]

Clean fix. The cross-publisher guard at publisher-db.ts:968 is load-bearing for catalog trust — lifting it would have been the wrong move; a distinct evidence value with its own writer is the right shape.

Things I checked

• ON CONFLICT clause in 488_catalog_fanout_authoritative_evidence.sql:70-74 and publisher-db.ts:314-318 matches the partial unique index idx_caa_unique_active (migration 440) exactly: (agent_url_canonical, COALESCE(property_rid::text,''), COALESCE(publisher_domain,''), evidence) WHERE deleted_at IS NULL. Both sites will fire the upsert path, not fail with no unique constraint matching ON CONFLICT specification.
• All 5 evidence-CASE sites in federated-index-db.ts (lines 164, 214, 280, 447, 1294) carry the new WHEN 'adagents_authoritative' THEN 'adagents_json' arm. The sort-priority CASE at line 1305 is over source (already-mapped column), not v.evidence, so it correctly stays untouched.
• Backfill join shape (488_catalog_fanout_authoritative_evidence.sql:65-69): publishers WHERE discovery_method='adagents_authoritative' ⋈ agent_publisher_authorizations WHERE source='adagents_json'. Right pair — recordChildPublisherFromManager sets the discovery_method, recordAgentFromAdagentsJson sets the legacy source. Bounded to ~thousands of cafemedia rows.
• recordCatalogFanoutAuthorization uses getClient() + try/finally for client release. Consistent with the writer pattern elsewhere in the file.
• Invalid agent_url silent-skip at publisher-db.ts:308 (if (!agentCanonical) return). Fail-closed beats fail-open. Test at catalog-fanout-projection.test.ts:131 exercises it.
• CHECK widening at 488_catalog_fanout_authoritative_evidence.sql:35-39 — DROP CONSTRAINT + ADD CONSTRAINT CHECK (...) takes ACCESS EXCLUSIVE during validation. At ~thousands of rows the window is sub-second. Acceptable.
• 6 integration cases cover row shape, idempotent re-write, coexistence with adagents_json for the same (agent, publisher) pair, invalid-url silent-skip, canonicalization, and view round-trip.

Follow-ups (non-blocking)

• Redundant BEGIN; ... COMMIT; in 488_catalog_fanout_authoritative_evidence.sql:32,76 — migrate.ts:130,142 already wraps each migration in a transaction, so Postgres will log WARNING: there is already a transaction in progress. Migrations 486/487 do the same, so it's convention now; either drop the inner wrapper repo-wide or accept it.
• PR body claims "6 evidence-CASE statements" in federated-index-db.ts; actual count is 5 (line 1305 is over source). Doc nit.
• Two test-plan checkboxes unchecked — backfill query plan on production catalog and post-deploy partner-snapshot probe. Both are reviewer/operator validation, not author-side gaps. The writer is covered by the integration suite; just don't skip the post-deploy probe.

Minor nits (non-blocking)

1. Raw vs canonical agent_url. publisher-db.ts:322 stores input.agentUrl raw (un-trimmed) in agent_url while writing agentCanonical into agent_url_canonical. Matches the existing writer pattern, so consistent — but the raw column is exposed to consumers; trimming the raw side would prevent " https://x.io " from round-tripping with leading whitespace.

LGTM. Follow-ups noted below.