[bokelley]

Summary

This PR fixes the three real Python bugs that the TS agent's parity notes on #1772 and #1773 pointed to (decision 4). I reproduced each one with a test that fails on next before applying the fix.

#1772: the re-grant chain always read new.

• While authority is lost, the account is missing from list_accounts. The snapshot acknowledged during the loss therefore had no authorizations entry for it, and account.regranted compared against nothing.
• I reproduced it against the managed seller: a re-grant within retention kept authorized_at = 2026-10-01T00:00:00Z and still emitted chain: "new".
• poll() now carries the previous snapshot's authorization forward for accounts that are in unauthorized and not listed now. This is TS's rememberAuthorizations.
• The snapshot is the durable source, so the fix also holds for serverless clients that start fresh on each poll.

#1773: adjustments were never offered.

• apply_receipt_policy reached the adjustments loop only after the "already receipted" continue. An adjustment corrects a revision after the buyer has receipted it, so the hook never saw one.
• Pending adjustments are now offered before that check, as in TS. An adjustment is pending when it has no adjustment_receipts entry.

#1773: rejected_by_buyer.

• Only an accepted receipt clears MISSING_MATCHING_CONSUMER_RECEIPT. A rejected billing revision therefore kept reading receipt: owed_by_buyer / responsible_party: buyer forever, even though the policy never resends a rejection.
• Such a period now reads receipt: {"status": "rejected_by_buyer"} with responsible_party: "seller", which corrects and re-delivers. This is TS's receiptStateOf, keyed on the obligation and the current revision.

The other #1772/#1773 notes needed no change: event ids and receipt ids are already equal across SDKs.

Docs: the buyer guide's status table and receipts rules, and the receipts module docstring.

Test plan

• test_losing_and_regaining_authority_are_events now asserts chain: "continued". The new test test_a_fresh_client_reads_the_chain_from_the_saved_snapshot uses a new client per poll over one store. Both fail on next.
• test_pending_adjustments_are_offered_on_an_already_receipted_revision: only the unanswered adjustment is offered and deferred, and nothing is re-receipted. It fails on next.
• test_a_rejected_revision_is_the_sellers_to_correct_not_owed_by_the_buyer: the receipt state goes owed_by_buyer, then rejected_by_buyer, and is None once accepted.
• The reconcile, events, receipts, TS-parity and test-kit suites pass. ruff and mypy pass on adcp.reporting.subscriptions.

🤖 Generated with Claude Code

[aao-secretariat]

Ladon verdict: Approve

Approve — clean three-bug parity fix against the TS buyer-reporting ports, no blocking or Medium findings.

Checked:

• Re-grant chain fix (_remember_authorizations) correctly carries prior authorization forward for currently-unauthorized accounts and still reads new outside retention.
• Adjustment reorder offers pending adjustments on already-receipted revisions, filtered by an answered set.
• _Obligation.receipt state machine routes rejected revisions to responsible_party: seller while keeping accepted → None.
• receipt_owed → receipt rename is private with both call sites updated; rejected_by_buyer is an internal dict value (no published-enum/wire-pathway concern).
• Each fix ships a reproducing test.

No high-risk flag, no gated paths, no author-team gate, no prior decision. Zero findings → decision table falls through to row 9.