[Patel230]

Summary

OMN-19968 asks that writers which run on a developer machine produce the same rows on SQLite and on Postgres. #3067 proved it for two writers. This adds the same proof for the other two writers that have SQLite tables, usage_by_model_day and metering_summary, so a local run and a lab run cannot silently disagree on the Usage page and the Overview savings figures.

What changed

One file, tests/test_omn19968_writer_store_parity.py (+154/-1). Two new cases, test_store_neutral_rows_equal_on_sqlite_and_postgres[usage_by_model_day] and [metering_summary]. Each runs the same store-neutral function once on the real SqliteDatabaseAdapter and once on the real PostgresSyncProjectionAdapter over the same fixtures (apply_usage_call over the ten-call fixture of tests/test_omn19978_usage_by_model_day.py; store_rows over the fold of a three-record request), compares the normalized rows, and asserts no Postgres-only SQL ($n, ::type, jsonb) is seen on the SQLite path. The node migration is rewritten from public. into the throwaway schema so the proof never touches a shared schema.

One helper fix in the same file: the normalizer handled datetime but not date, so a Postgres DATE column and SQLite's ISO text compared unequal. That was the first red run, a gap in the test helper and not a row difference.

No source file changes.

How it was verified

• uv run pytest tests/test_omn19968_writer_store_parity.py -v at the branch head: 8 passed, 0 skipped (6 before). PostgreSQL 16 was a disposable local cluster built by the test's own fixture.
• First run, before the normalizer fix: 7 passed, 1 failed (usage_by_model_day, usage_day string versus date).
• Mutation proof on src/omnimarket/projection/sqlite_database.py _encode, worktree only, restored after each. Baseline: 8 passed. Rounding Decimal to two places: 1 failed (usage_by_model_day). Truncating strings over 10 characters: 4 failed (delegation_events, llm_call_metrics, metering_summary, usage_by_model_day). After restoring: git diff on the source file empty, 8 passed. Rounding missed llm_call_metrics, whose cost is a float that survives rounding at the compared precision.
• ruff format and ruff check clean; every pre-commit hook passed; the commit is signed.
• wf pr selfcheck: branch name, no AI attribution, all commits signed, 155 changed lines, no core.hooksPath override all pass.
• Overlap: wf pr checked this one changed path against 6 open omnimarket PRs and found no overlap. My own earlier read listed 8 open PRs and none touched the file; the two counts differ and I did not reconcile why. feat(OMN-20154): provider quota state is a durable projection that routing and the judge read, every metered call is observed, and 1302 is a capacity cooldown #3134 lists 99 files, at the 100-file cap, so its list may be truncated.

Not proven here: the async Postgres runtime writers (asyncpg pool, advisory lock) are covered only by their own real-Postgres tests, which skip without INTEGRATION_POSTGRES_PASSWORD and which I did not run. The Postgres leg ran on a local cluster, not on the lab host.

Failure paths

n/a: this change adds tests only and writes no receipt, verdict or report.

Open defects

none

Local gates

• No gate was bypassed — no --no-verify, no SKIP=, no --no-gpg-sign, no core.hooksPath override. Every pre-commit and pre-push hook ran.
• Any hook that failed was fixed at the input, not worked around.

Every downstream proof — receipts, dod_verify, the evidence chain — assumes the local gates ran,
and the PR body is the only place that assumption is written down. core.hooksPath is the silent
case (OMN-16725): in a worktree .git is a file, the override resolves to nothing, git finds zero
hooks, and the commit succeeds with no warning.

Not in this change

• node_projection_tenant_credentials has no SQLite table and still uses Postgres-only SQL. It is not covered here; the proposal to carry it onto OMN-19985 is with Jonah on OMN-19968, not decided.
• The ticket's acceptance criteria are not closed by this PR. The evidence companion and any criterion binding come after merge.
• No change to the writers themselves, and no change to the async runtime writers.

Evidence-Ticket: OMN-19968
Evidence-Source: OCC#11990

[github-actions]

✅ Architectural Review — PASSED

Errors: 0
Warnings: 0

What this checks

Rule	Description
ARCH-TOPIC-001	No hardcoded Kafka topic strings in handler code
ARCH-DI-001	No event_bus=None bypass
ARCH-DI-002	No direct Handler instantiation outside workflow_runner/adapters
ARCH-DI-003	No reinvented DI containers
ARCH-TOPIC-002	contract.yaml topics follow `onex.{cmd

---

No architectural violations found.

Static architectural lint — no model inference (OMN-14176).

[github-actions]

✅ Hostile Reviewer — PASSED

Blocking findings (critical/error): 0
Total findings: 10
Models succeeded: qwen3-review,gpt-oss-review

---

Gate semantics

Verdict	Meaning	Blocks merge?
passed	>=2 models succeeded, no critical/error findings	No
blocked	CRITICAL or ERROR findings found	Yes
degraded	Fewer than 2 models succeeded (infra unavailable/timeout)	Yes (OMN-15110)

Powered by omniintelligence.review_pairing.cli_review — node-based adversarial review via HandlerLlmCliSubprocess (OMN-8524)

[Patel230]

No reviewer left a comment here. The hostile reviewer's 10 findings were all below quorum and posted no threads; they were style points such as the inline __import__("datetime") calls, and no change was made for them.

Merged as 66af321 on 2026-10-01 for OMN-19968. It changed two files: the store parity test (tests/test_omn19968_writer_store_parity.py) and an import-order fix in benchmarks/delegation_false_pass_growth/readout.py. Every check on the final head passed or was skipped.

Left on the ticket: omnimarket#3183, the llm-cost writer fix, open with every required check green and waiting for review and merge.

Thank you!