Workflow Tune Verify

Purpose

Fix the verification responsibility of the Tune stage candidates, and select the adopted candidate by objective metrics. This flow is part of an optional flow separated from the core workflow.

Scope

• the trial-result evaluation of a candidate lowering-plan (target_lowering_plan) variant
• the finalization of the best impl and the re-tuning judgment

Requirements

• Exclude a physics fail candidate from the performance-evaluation target.
• Do not adopt a candidate whose quality-comparison result is a failure.
• The performance evaluation uses the statistics of perf.json, and ranks by the objective function.
• Handle the re-measurement result of the same point, and make a noise-robust judgment.
• Make it a required condition that the adopted candidate is traceable in trial_meta.json and lineage.json.
• The workflow mode uses ATMOFAB_WORKFLOW_EXEC_MODE as the canonical source, and applies dev when unset.
• In dev mode, it is a Tune fail the moment issue_severity=major|critical is detected, and treating it as a minor exception is forbidden.

Operations Rules

1. Keep only candidates whose verdict and aggregate_verdict are pass as adoption candidates.
2. Among the adoption candidates, fix the variant lowering plan with the maximum objective function as the best impl.
3. Under a new-architecture or new-compiler condition, run the re-tuning judgment.
4. Save the judgment basis in the tuning-family metadata, and hand it to regression monitoring.
5. When it fail in dev mode, record the basis needed to create failure_analysis.json (the failure reason, the target trial, the judgment evidence).

Decision Criteria

• The adopted candidate satisfies the required physics and quality conditions.
• The performance ranking can be reproduced from the saved values of perf.json.
• The finalization basis of the best impl is traceable.