[shuqi2077]

Summary

Addresses #1517 by extending the existing RMSNorm autotune example from printing a winner to actually consuming persisted winners in application dispatch.

Changes

• Add separate tune RECORD_PATH and run RECORD_PATH modes. The run-only path never starts a tuning search.
• Persist winners for two exact shape buckets, including measured timing/sample provenance and tagged L2 keys.
• Build the current Workspace from source, architecture, assembler and search-space identity; reload through Record::load_verified and recompute keys using the same kernel builder used for execution.
• Select an eligible recorded configuration by input shape, with a fixed fallback for a missing/rejected record or a missing bucket. Verification warnings are displayed and conservatively fall back as an explicit example policy.
• Check candidate and dispatched outputs against a CPU RMSNorm reference. The no-argument demo exercises save/reload, both recorded buckets and an untuned fallback bucket, then removes its temporary record.
• Document the complete workflow. No runtime API changes, new dependencies or imported RUDA implementation.

Review status

Draft: the required full Tile validation is still outstanding; this PR does not claim GPU validation or performance gains.

Contribution

Prepared with OpenAI Codex assistance. The commit includes DCO sign-off and a verified GitHub signature. Existing source attribution is preserved. No generated tuning records or internal development reports are included.

[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.