[anle2024]

Design discussion: #139 (review)

Summary

Add P1 as the separate CPU sampling change requested in #139. Five complete operations use the frozen commerce/100M graph and 300000 seed evaluations each. The candidate must preserve counts, fetched rows, and sampled-node membership. Its median time must meet the reviewed reference budget. P1 does not fit or load a model.

The manual weekly candidate run builds one wheel. A separate P1 job installs that wheel after the eleven-case GPU jobs finish. The shared host lock covers the native block. P1 uploads its evidence, removes temporary work, and has a 120-minute timeout. There is no cron trigger.

Raw measurements and the approved reference are separate records. The reference binds the same data and work, the original measurement checksums, and the reviewed timing policy. A run cannot approve its own reference. Candidate dispatch now requires a p1 entry in KUMO_ENGINE_WEEKLY_CONFIG. Configure its exact input and reference checksums after review. Missing bindings stop configuration before native work.

This PR is based on #139. B9 is #150, based on this branch so the jobs share one workflow and run in order.

Testing Done

40 local CPU checks pass. Workflow configuration, candidate-wheel binding, helper compatibility, complete work, changed membership, timing drift, and reference lineage are covered. Actionlint, Python style, and documentation checks pass.

Three direct native calibration blocks completed all fifteen operations with the published 1.0.0 wheel and no model work. Their medians range from 89.67 to 89.97 seconds. Counts and membership match across all operations. The public calibration record contains the measurements and checksums. It proposes a 103.67-second median limit, with ten percent plus five seconds of operational headroom. This allowance still needs review.

Pending qualification

Keep this PR draft until the reference and timing policy are approved and one manual candidate run passes. These direct baseline runs do not qualify the full GitHub workflow or a different candidate wheel. No GitHub workflow was dispatched. Automatic push/PR CI is skipped to preserve the no-training scope.

Change map

Maintained change map.

[anle2024]

Requested change: Split P1 CPU sampling gates from the eleven-case runner in #139.

Existing behavior: PR139 runs the eleven model cases. It has no P1 sampling job.

Smallest sufficient change: Add one repository-owned sampling runner and one sequential manual weekly job that uses the shared wheel and a separately reviewed reference.

Focused verification: 40 local CPU checks and fifteen native sampling operations pass. No fit, GitHub dispatch, cron activation, or full candidate workflow ran.

Stopping condition: Keep the change draft until a maintainer approves the measured reference and timing policy and one manual candidate run passes.

Change map

File	Function / scope	What changed	Why
.github/workflows/kumo-engine-weekly.yml	Configuration and p1 job	Bind P1 inputs and reference, reuse the candidate wheel, serialize native work, and upload before cleanup.	Give P1 a separate result and timeout on manual weekly runs.
docs/development/engine-weekly-ci.md	Auxiliary checks	Link the separate P1 job.	Keep the weekly contract accurate.
docs/development/engine-weekly-p1-calibration.json	Native baseline measurements	Record fifteen operation timings, exact work checksums, and the proposed reference policy.	Expose measured numbers without private host links or automatic approval.
docs/development/engine-weekly-p1.md	P1 contract and approval	Define the work, timing formula, reference lineage, configuration, and evidence lifecycle.	Make the gate and approval process reviewable.
packages/kumo-engine/scripts/weekly_p1.py	inputs, sample_worker, run_locked, reference_operations	Run full native sampling with feature fetch, compare counts and membership, and enforce the reviewed median budget.	Detect changed sampling work and timing drift with the selected installed wheel.
packages/kumo-engine/test/test_weekly_p1.py	SamplingChecks	Reject incomplete work, changed membership, invalid timing, wrong installed bytes, and mismatched reference lineage.	Exercise the failures that must stop P1.
packages/kumo-engine/test/test_weekly_workflow.py	Workflow configuration and P1 launch	Execute the inline programs and check candidate-wheel, manifest, reference, and host-lock bindings.	Catch wiring errors before native sampling starts.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts