[bokelley]

Refs #1669

The SDK does not produce get_media_buy_delivery itself. I checked each AdCP 3.3 reporting rule against the SDK code that reads, derives, injects, validates or stores delivery and reporting data, and fixed only the real gaps.

Rule	What the SDK owns	Change
impressions optional on delivery rows	Generated row types (already optional). InlineReportingSource projects delivery into rows and control totals.	No code change was needed. _metrics already omits absent metrics, and _control_totals skips a metric that some rows lack. New tests pin both. No SDK reader derives cpm, ctr, completion_rate or viewable_rate.
Multi-qualifier metrics omit flat totals	Nothing. No SDK code builds buy, window or webhook totals.	Docs only.
Receivers ignore aggregated_totals in delivery webhooks	WebhookReceiver, ADCPClient.handle_webhook	Both already accept the field (the webhook result is open) and never read it. New tests pin this.
Property and installment-property rollups	comply_test_controller dispatch of simulate_delivery	The dispatcher now rejects property_delivery, installment_delivery and installment_property_delivery with INVALID_PARAMS when they come with delivery_date (the arrays are undated). It also rejects them when the store override does not declare the parameter. Before, it dropped those rows and reported success. Docstring updated for injected rows and optional impressions.
source.custody on report definitions	ReportingRevisionVerifier, which hashes and parses the installed definition bytes	New ReportingRevisionVerifier.custody property. A custody value outside the enum fails construction with UNSUPPORTED_VERIFICATION. The digest already covered the claim because raw bytes are hashed. The reference definition asset is unchanged, so its digest does not move.
Superseded snapshot compaction	Ledger retention and exact reads	Docs only. Compaction is not permitted under 3.3, and keeping the full chain stays conformant. The plan is in docs/reporting-release-notes.md.

No adcp-reporting-storage change was needed. Its row codecs are opaque JCS rows, and its definition column stores the URI and digest binding, not the definition document.

Tests

• tests/test_reporting_33_rules.py (new, 22 tests)
• make lint, make typecheck-all: pass
• pytest tests -n 8 -k "reporting or test_controller or webhook": 4942 passed, 4 failed. All 4 are subprocess import tests (test_lazy_types x2, test_reporting_staging_reuse, test_reporting_production_public). They pass when re-run on their own, so they appear to be load flakes. tests/test_server_dx.py and tests/test_account_mode_gate.py: 141 passed.

🤖 Generated with Claude Code

[aao-secretariat]

Ladon verdict: Approve

Approve — AdCP 3.3 reporting rules applied cleanly across three surfaces with no blocking or medium findings.

What I checked:

• high_risk: true is driven solely by src/adcp/server/test_controller.py (modified) matching src/adcp/server/**. Per the risk model, a modification with no medium-or-higher finding is presumed safe. The reviewer confirmed simulate_delivery dispatch correctly reads the bound override's real signature via getattr(store, scenario) + inspect.signature — _LegacyStore is refused, **kwargs store passes (test_controller.py:1413-1420) — and refuses undated injected rows rather than silently dropping and reporting success. That satisfies the DR-0032/DR-0033 pathway rule (Refused path, with a test).
• ReportingRevisionVerifier.custody fail-closes on malformed claims via _runtime() in __post_init__ (verification.py:96, 127-135), covered by the definition digest.
• Semver signal feat(reporting): is correct: additive property, no public export removed/renamed/retyped, no required↔optional flip.
• New 488-line test file pins real behavior (control-total omission at tests/test_reporting_33_rules.py:351; digest coverage at :522). The 4 disclosed test failures are documented subprocess-import load flakes on an unchanged path, not validation gaps.

Gates: gated_paths false; no no-auto-approve team match; no prior decision. Decision table rows 1–8 do not fire (no critical/high, no medium, high_risk modification with zero findings) → row 9 approve.