[Patel230]

Summary

node_projection_llm_cost now writes the shared usage-source vocabulary (measured / estimated / unknown) on both stores. It used to write the retired API / ESTIMATED / MISSING, which every database omnibase_infra has migrated refuses. That is why this ticket's own lab step failed (OMN-19968, Amendment 2, AC8–AC11). Merges after OmniNode-ai/omnibase_infra#4420, which vendors this PR's migration.

What changed

• src/omnimarket/nodes/node_projection_llm_cost/handlers/row_llm_call_metrics.py: _usage_source() maps provider-reported usage to measured, ESTIMATED to estimated, and MISSING / UNKNOWN / absent / unrecognised to unknown. The retired labels are folded in case-insensitively. usage_is_estimated derives from it. The docstring that still described the old mapping is corrected.
• src/omnimarket/projection/sqlite_database.py: usage_source defaults to 'unknown', and existing SQLite rows carrying a retired label are moved onto the shared vocabulary in place.
• src/omnimarket/nodes/node_projection_llm_cost/migrations/0003_usage_source_shared_vocabulary.sql (new): renames an omnimarket-created usage_source_type's labels in place (RENAME VALUE, no table rewrite) and sets the default to 'unknown'. The type is resolved through search_path (to_regtype), because 0001's typname guard matched a same-named type in any schema. That is how an infra-migrated database ended up rejecting this writer. On a database that already carries the shared vocabulary it changes nothing.
• Tests and fixtures in 7 files now use the shared values. New tests/unit/test_omn19968_usage_source_vocabulary.py (17 cases) pins the mapping, the SQLite schema default and reconciliation, the migration's in-place renames and the migration number's uniqueness.

Grepped for every remaining use of the retired values (spec approach step 2). Left on purpose: the reconcile CASE in sqlite_database.py, the test inputs that feed API/ESTIMATED to prove the mapping, a comment in savings migration 087 that quotes a producer's wire label, and EnumUsageSource._missing_ (src/omnimarket/enums/enum_usage_source.py, from OMN-10382), whose legacy-alias table maps the same old labels to the same shared values this writer now uses. No reader treats 'API' as a meaning of its own.

How it was verified

Commit 6c26fd7ec (signed, G), on dev c603479d9.

• RED then GREEN, same base: the 8 changed test files fail on unfixed dev (35 failed, 92 passed, including 17/17 of the new file) and pass on this branch (127 passed, re-run on 6c26fd7ec).

• The tests can fail (AC11): two mutations in the worktree only, each restored byte-identical (git hash-object checked):

  • mapping api back to estimated → 5 tests fail;
  • the SQLite reconcile sending MISSING to estimated → 1 test fails.

• Related suites, dev versus branch at the same sha: every test file that touches llm_call_metrics, usage_source, sqlite_database or node_projection_llm_cost (60 files) on dev ce2f9cfba and on the branch on ce2f9cfba: dev 1015 passed, branch 1032 passed (+17, the new file). The same single failure on both, tests/integration/test_cost_event_publisher_kafka.py::test_idempotency_key_is_stable_on_reprocess, is environmental: no Kafka at localhost:19092.

• Lab, the ticket's own step (AC10): tests/test_omn19968_writer_store_parity.py on the lakshman lane's PostgreSQL 16 on .201, in a throwaway container on the lane network (2026-10-01T10:51:49Z): 8 passed, 0 failed. The same file with the same method on unfixed dev at 07:46Z: 6 passed, 2 failed (invalid input value for enum usage_source_type: "MISSING").

• Migration 0003 on real PostgreSQL 16.15, lakshman lane (AC9), 10:08:06Z, throwaway databases dropped afterwards: 14 of 14 assertions in 4 starting states:

  • omnimarket-built with stored rows: renamed, rows kept, default 'unknown', 'API' refused afterwards, idempotent;
  • shared vocabulary already present: unchanged;
  • no type at all: clean no-op;
  • same-named type in another schema: untouched.

  Mutation control: removing the MISSING rename fails 3 assertions.

• mypy src/omnimarket/ --strict: no issues in 3966 files. ruff check and ruff format --check: clean.

• scripts/ci/check_wire_compatibility.py over the changed files: WIRE_COMPAT_NO_WIRE_CHANGE. Positive control: a wire model file gives WIRE_COMPAT_PASS.

• scripts/check_release_identity.py --base origin/dev: 0.4.272 ahead of the published 0.4.271, so no bump is needed.

• check_node_metadata_dependencies.py: PASS. node_runtime_sweep --import-check: exit 0.

Update 2026-10-03, the pin (commit 0e413dfbe, signed G): the Sibling Release Compatibility Gate (OMN-20381) refused migration 0003 because the pin, 0.38.61, is a release without it. The floor is now omnibase-infra>=0.38.63,<0.39.0 in both dependencies and override-dependencies; omnibase_infra#4420 (98620a57) is first released in v0.38.63. uv.lock was refreshed for omnibase-infra only with uv 0.12.22, the version CI installs: revision stays 5, 10 lines change, no other package moves; uv lock --check passes.

• The gate's own comparison run locally against omnibase_infra tags: v0.38.61 lacks docker/migrations/forward/nodes/node_projection_llm_cost/0003_usage_source_shared_vocabulary.sql; v0.38.63 carries it byte-identical.
• Related suites (-k "llm_cost or llm_call or usage_source or metering or parity", -m "not kafka and not integration") on dev 76e1673d0 and on this branch merged with it: dev 795 passed, 5 skipped, 0 failed; branch 816 passed, 1 skipped, 0 failed.
• AC11's mutation re-run on the new head: removing the "api" synonym fails 3 of the 17 vocabulary cases; restored, 17 pass.
• Two merge commits on this branch, c47936adc and 94eb9c2a1, were made by GitHub's update-branch and are signed by GitHub (verified: true on the commits API); the local check cannot verify GitHub's key.

Overlap-Reviewed: #3172 — both edit tests/test_projection_handlers.py in separate hunks: #3172 adds model_cloud_baseline to two delegate-skill fixtures (near lines 488 and 847); this changes one expected usage_source value in TestLlmCostHandler (near line 208) from 'MISSING' to 'unknown'. No shared lines; each is needed for its own change and they merge in either order.

Overlap-Reviewed: #3309 — it raises the same omnibase-infra floor to 0.38.63 for the repo (OMN-17427). This PR needs that floor for its own migration and cannot wait on #3309, which is red on other checks; the two produce the same pin, so whichever merges second takes either side of the pyproject and uv.lock hunks.
Overlap-Reviewed: #3308 — it bumps omnimarket's own version line in pyproject.toml and uv.lock (0.4.284 to 0.4.285); this PR changes the omnibase-infra dependency lines. Different lines, both needed.
Overlap-Reviewed: #3245 — it bumps omnimarket's own version line (0.4.282 to 0.4.283); same reasoning as #3308.
Overlap-Reviewed: #3293 — it adds nullable finish_reason and truncated to delegation_events in sqlite_database.py; this PR changes only llm_call_metrics there. Different tables, both needed.

Failure paths

n/a: this change 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.

The pre-commit dependency health gate first failed because it ran under the system Python 3.9 (datetime.UTC missing). It passed once the worktree's own venv was first on PATH, as the workflow rules require; no hook was skipped.

Not in this change

• node_projection_tenant_credentials's SQLite path. Per the ticket's Amendment 1 it is proposed to move to OMN-19985, pending Jonah. AC7's reply on the ticket waits for Lakshman's YES.
• savings_estimates.usage_source and any other table's vocabulary (out of scope in the plan).
• Merging; that is Jonah's.

Evidence-Ticket: OMN-19968
Evidence-Source: OCC#12273

[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)

[onexbot-occ-writer]

OCC autobind rebound this PR's evidence-source stamp line to OCC#12273.

The receipt gate's own eligibility validator, run against the change-control tree the gate pins for OCC#12273, returned eligible for head 94eb9c2a1f5eff16adb8ffb5fa0fa203eed09f58. The body now carries exactly one stamp line. Displaced: OCC#12139.

Reported by occ_companion_emitter (OMN-18853).

[Patel230]

Answers to the two hostile-review notes on this head (job 111263134881, both info, from gpt-oss-review):

1. _usage_source and EnumUsageSource inputs (row_llm_call_metrics.py): not a defect. EnumUsageSource is a StrEnum (enum_usage_source.py:12), so str(member) is its plain value, and an enum or a string maps to the same member at row_llm_call_metrics.py:64. The projection handler also passes event.model_dump(mode="json") (handler_projection_llm_cost.py:126), so the value arriving there is already a string. No change.
2. SQLite default not migrated for existing databases (sqlite_database.py): the rows are migrated. _reconcile_legacy_llm_call_metrics runs on every open (sqlite_database.py:288) and rewrites API, ESTIMATED and MISSING to measured, estimated and unknown (:309-318), and running it again changes nothing. Only the column default of a table created before this change stays 'MISSING': SQLite cannot alter a column default, and :211 applies to new tables. Nothing relies on that default, because build_llm_call_metrics_row always sets usage_source (row_llm_call_metrics.py:198) and the writer stores the whole row (handler_projection_llm_cost.py:129-140). No change.

Thank you!

[jonahgabriel]

Thanks, @Patel230 (Lakshman), for omnimarket#3183 on ticket OMN-19968, which was squash-merged as 5de6e9d on 2026-10-04T23:13:51Z by the landing lane under the operator ruling of 2026-10-04T23:07:50Z that overrode the report-only landing policy for the collaborator pull requests listed in the standups ask. We really appreciate your contribution!