[dbccpage]

Requested by Jeremy · project thread

Follow-up to the adversarial review of #47 (already merged). Stacked on #51; only the last commit is new.

Review item	Change	Regression test
P1-1 receipt budget	Every transaction costs max(1, len(change)) event units in both peek_many and apply_many, because an empty transaction still yields an empty_transaction receipt per binding.	test_empty_transactions_count_against_the_receipt_budget: two empty txs, max_events=1, max_outputs equal to the binding count. Fails on parent (invalid_generation_state); here it publishes the admissible prefix, then the rest.
P1-2 generation may not fit	apply_many publishes the longest prefix that fits. It first cuts to the event budget, then halves on state_limit / state_bytes_limit / invalid_generation_state, before anything commits. It acks only that prefix and carries freshness only if the whole batch went out. A single transaction that does not fit is refused as before.	test_an_oversized_generation_publishes_the_largest_prefix_that_fits, test_the_group_is_held_to_max_events (now a prefix, not a refusal), test_a_single_transaction_over_max_events_is_still_refused
P1-3 server-side decode bound	A coalesced peek passes upto_lsn = min(frontier, durable_offset + max_bytes), so PostgreSQL decodes at most one max_bytes window of WAL per poll, not max_coalesce × max_bytes. If that window holds no commit, the next transaction is peeked alone up to the frontier (the pre-#47 behavior).	Native test_production_coalesced_peek_decodes_a_bounded_wal_window: 3000 commits, max_bytes=128 KiB, max_coalesce=1024. Every coalesced peek's upto is within the window, the window binds, oversized generations fall back to prefixes, and the oracle matches.
P2-4 profiler	Instruments peek_many / apply_many. Reports transactions_published, generations_published, transactions_per_s from published transactions, and connections per transaction. Contention evidence adds fraction_of_commits_observed.	Phase profile run: 4 polls, 4 generations, 200 transactions.
P2-5 CLI	Optional max_coalesce (1–1024, default 64; 1 = strict per-transaction), passed to PostgresWorker and printed by check. Documented in docs/production_profile.md.	tests/test_production_cli_config.py
P2-6 overlap	apply_many requires transactions[0][1] > previous_offset once duplicates are resolved.	test_a_batch_overlapping_the_durable_offset_is_refused

Not done here: the full differential suite across extremum/TopK/joins/set operations, and testing under constrained PostgreSQL memory. Both are worth a separate PR.

Local evidence (PG 16.15 + wal2json + Redis 7): the 4 new runtime tests fail on the parent branch and pass here. Production unit tests: 170 passed. Native suite: 60 passed. black and mypy are clean.

---

Generated by Claude Code