[jonahgabriel]

What

New node node_projection_usage_by_model_day (OMN-19978, plan task T3.6): tokens in, tokens out, cost and call count per (tenant, UTC day, model id), from onex.evt.omniintelligence.llm-call-completed.v1. Bus-backed exposure onex.snapshot.projection.usage-by-model-day.v1 (cursor projection_cursor, key columns tenant_id, usage_day, model_id).

• Pure definition-B fold HandlerProjectionUsageByModelDay: UTC day derived from the event timestamp, never the clock; a missing timestamp is refused (poison to DLQ). Missing tenant falls to the house tenant.
• Store-neutral apply_usage_call(delta, db) over upsert/query only: ledger row keyed on call id, then RECOUNT of the key (never increment), so a replay writes nothing. Runs on the real SQLite adapter (DDL added beside the others in sqlite_database.py).
• Effect writer UsageByModelDayProjectionWriter (in-process dispatch, only the writer is routed): one transaction, advisory lock on the key, insert-if-new, recount, snapshot delta.
• Tenant domain, not internal: both tables carry tenant_id, so they live in public with row-level security (0001), like node_projection_cost_summary.

Reuse

node_projection_cost_summary does NOT carry this grain (its key is a 24h/7d/30d window enum with no model or day), so it is not reusable here. Reused: the node_projection_topic_activity bus-backed contract and writer shape, SqliteDatabaseAdapter, HOUSE_TENANT_SLUG, cost_summary's RLS migration shape.

Evidence (dod_evidence)

• AC1: uv run pytest tests/test_omn19978_usage_by_model_day.py -- ten calls, two days, two models give four rows; token, cost and call-count sums equal the call rows'; replay changes nothing (19 passed with the writer tests).
• Lab readback (real Postgres, dev-lane database on the .201 lab host, throwaway schema): tests/test_omn19978_usage_by_model_day_real_postgres.py ran the writer's own insert and recount statements for the same ten calls: 1 passed; four rows, sums equal, replay left rows and cursors unchanged.
• Ratchets moved with the node: relation inventory (+2 tables), exposure and cursor counts, node and entry-point counts, golden chain.

Not proven here (stated plainly)

• AC2 (the rendered Usage page equals the sum of the day's call rows) is NOT proven: no runtime deploy, snapshot-topic read or page render was done in this lane. It needs the deployed writer plus the page reader; follow-up on the ticket.
• The SQLite exposure read side is the mode 1 reader (T2.3), not this PR. The SQLite tables carry no projection_cursor.
• omnibase_infra vendoring of the new migrations (vendor-parity gate) is a companion step if the gate demands it.
• Depends on OMN-19968 (writer store parity) only in spirit: this node has its own store-neutral path and does not touch its files except sqlite_database.py DDL (adjacent hunk, may need a trivial rebase).

Delegation

Codex drafted the node from the RED test; the lab (onex delegate --task-type code_review, Qwen3.8-27B, whole files) reviewed it: three findings, all discarded with reasons (single-process local path, sequential by design; the Postgres path holds the advisory lock; a primary key column is already NOT NULL).

Evidence-Ticket: OMN-19978
Evidence-Source: OCC#11871

[github-actions]

⚠️ Architectural Review — WARNINGS

Errors: 0
Warnings: 1

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

---

ARCH-DI-002: Direct Handler instantiation (warning)\n\n\nsrc/omnimarket/nodes/node_projection_usage_by_model_day/handlers/handler_usage_by_model_day_writer.py:86: self._derive = HandlerProjectionUsageByModelDay()\n\n\nHandler classes should be wired only in workflow_runner.py or adapter modules. Direct instantiation elsewhere bypasses the contract system.\n\n

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

[github-actions]

✅ Hostile Reviewer — PASSED

Blocking findings (critical/error): 0
Total findings: 18
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)

[onexbot-occ-writer]

OCC autobind rebound this PR's evidence-source stamp line to OCC#11779.

The receipt gate's own eligibility validator, run against the change-control tree the gate pins for OCC#11779, returned eligible for head 179e4a36d25796e29b77c149b1ce70139ab1756a. The body now carries exactly one stamp line. Displaced: OCC#11772.

Reported by occ_companion_emitter (OMN-18853).