[Patel230]

Summary

Six snapshot exposures answer 503 projection_column_missing on the .201 dev lane since omnimarket#3208: delegation.summary, delegation.model-routing, delegation.quality-gate, delegation.token-usage, delegation.savings and cost.savings-overview. On 10-01 all six answered 200 and matched their writers. That is the Overview page's delegation and savings data. Found in the OMN-19971 readback of 2026-10-02 06:01Z (ticket comment bee62174); it also re-fires the catalogue check that no exposure marked ok answers 503.

Cause. #3208 added a latest-row-per-key read that selects, orders and compares on every declared key_columns entry. These six declare key_columns: [snapshot_grain, tenant_id]. snapshot_grain is the constant half of the bus compaction key, minted by the republish query, and deliberately not a column of their views (node_projection_delegation/contract.yaml, the snapshot_grain comment). Postgres raises UndefinedColumnError, and the reader maps it to 503.

What changed

• src/omnimarket/projection/table_reader.py:
  • TableRowSource reads the relation's live column names once per relation from pg_attribute, with the relation as a bound value (to_regclass($1)), cached beside the cursor-type cache. It reads them only for an exposure that has key columns and a recency column.
  • New physical_key_columns(): the declared key columns the relation carries, in declared order. build_window_query takes relation_columns and keys the latest-row-per-key read on those. With none left, the per-key read is skipped.
  • A relation that does not resolve gives no columns, so every declared key is kept and the read itself still names the missing table.
• No contract changes: dropping snapshot_grain from key_columns would break bus compaction (one live record per tenant per aggregate).
• Version 0.4.279 (pyproject.toml, and only the omnimarket version line in uv.lock).

How it was verified

Failure modes written before the implementation, each with a test:

• F1: a virtual key column reaches the SQL.
• F2: a key column the relation does carry is dropped, collapsing distinct keys (the consumer-flow defect fix(OMN-20327): fix projection API cursor walks and lane-health contract #3208 fixed).
• F3: every key column is virtual and the per-key read runs with no keys.
• F4: a declared non-key column the relation lacks stops being refused by name.
• F5: the column lookup is interpolated or repeated, or a missing relation stops being refused by name.

Tests:

• tests/unit/projection/test_omn19971_virtual_key_columns.py (4): SQL shape for F1–F3, and the bound, cached lookup (F5).
• tests/integration/projection/test_omn19971_virtual_key_real_postgres.py (5), on a real disposable PostgreSQL: a per-tenant view with no snapshot_grain (F1), consumer-flow's shape with one busy and two quiet keys (F2), a virtual-only key (F3), a missing non-key column (F4), a missing relation (F5).
• Red before: with dev's table_reader.py (750808959), F1 and F3 fail and F2, F4 and F5 pass. The last three are guard rails, and they pass before and after.
• Green after: 9 passed; tests/unit/projection, tests/unit/nodes/node_projection_read_effect, tests/integration/nodes/node_projection_read_effect and the new files give 443 passed, 5 skipped. mypy --strict on table_reader.py, ruff check and ruff format --check are clean.

Lab first, on the lakshman lane on .201 (read-only, 2026-10-02T07:41:16Z):

• For each of the six exposures, build_window_query built the SQL from the real contracts. relation_columns came from the lab views' own pg_attribute, and the tenant was 820272f9-…. Each query ran in a READ ONLY transaction on omnibase-infra-lakshman-postgres.
• Dev's SQL: 6 of 6 fail with ERROR: column "snapshot_grain" does not exist.
• This branch's SQL: 6 of 6 return 1 row, equal to each view's row count for that tenant.
• Not covered here: the read through a rebuilt lakshman-lane projection API, and the 69-exposure readback on the shared dev lane. The second needs this merged and the dev lane redeployed; it then re-proves that no exposure marked ok answers 503.

Failure paths

n/a: this change writes no verdict, receipt or report. It changes which columns one read keys on.

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.

Not in this change

• The cursor-walk fixes in fix(OMN-20327): fix projection API cursor walks and lane-health contract #3208, which stay as they are (F2 guards them).
• The SQLite source, which never reads key_columns.

Overlap-Reviewed: #2965 — both touch pyproject.toml and uv.lock: #2965 raises the omnibase-core floor to 0.47.24 and removes its git-rev override; this changes only omnimarket's own version line (0.4.278 to 0.4.279). Different lines and different purposes, so each is needed; whichever merges second takes the next version.

Evidence-Ticket: OMN-19971
Evidence-Source: OCC#12295

[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: 12
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]

The CodeQL note on tests/integration/projection/test_omn19971_virtual_key_real_postgres.py (unused _FIXTURES, 2026-10-02 08:08Z) was fixed before merge in db029a5c9: the test now requests the disposable PostgreSQL fixture by name, so the unused global is gone.

Merged as efc25b3 on 2026-10-02. It fixes the 503 projection_column_missing on the six snapshot exposures that declare snapshot_grain; every test split and CodeQL passed on the final head.

Left on OMN-19971: AC1 is still not met on the dev lane because of two other problems read on 2026-10-03 (intermittent 503 timeouts on three exposures, and a decisions window serving old rows).

Thank you!