Send point.yaml and system_power.json, with derived power, on runs create

[arav-agarwal2]

Pairs with mlcommons/mlperf-endpoints-api#28, which adds runs.point_config and runs.system_power.

Today only system_desc.json, config.yaml and the result summary reach the database. point.yaml and system_power.json travel only inside the GCS archives, so the visualizer can't see a point's offline mode, its nodes_used, or any power figure.

runs create now sends:

• point_config: point.yaml, as parsed.

• system_power: when the run folder has system_power.json, the descriptor unchanged, plus a derived object:

  key	meaning
  provisioned_power_kw	E.5 total, or null if E.7 rejects the descriptor
  point_power_kw	§4.5.3, scaled by the point's nodes_used
  system_tps_per_kw	system_tps / point_power_kw, to one decimal (§4.5.2)
  mlc_estimated_power	true if any Appendix D default was used, or if the submitter set the flag
  estimated, problems	what the computation relied on, and why a figure is null

New submission_checker.power_summary.derive_point_power:

• It runs SystemPower.compute() and PowerComputation.point_power_kw(), so a stored figure can never disagree with what compliance accepts.
• It works on parsed contents rather than paths, so the review-repo sync job can import it and recompute when a PR edits these files.
• It never raises. If a descriptor or nodes_used is something the checker would reject, the affected figure is null and problems says why.
• The file in the archive is untouched; derived exists only in the database copy.

Compatibility: an API without #28 ignores the new fields, because FastAPI drops unknown body keys, so this can ship in either order.

Testing:

• New tests/submission_checker/test_power_summary.py covers: fully engaged, scaled by nodes_used, Appendix D default, submitter flag, over-provisioned and unknown-ensemble nodes_used, invalid descriptor, missing throughput, no mutation.
• Parser tests cover the passthrough and the derived figures against the fixture (14.7 kW).
• Ruff and mypy are clean. The full suite passes except test_builder.py::test_a_separate_accuracy_run_still_wins, which fails on main too, on local Python 3.10.8 (tarfile.extractall(filter=)); CI uses a newer patch release.
• Docs: docs/endpoints-cli/usage/runs.md.

🤖 Generated with Claude Code

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅