[power data] Complete energy coverage for promoted B200, H200, and GB200 cohorts

[jasonqinzhou]

Context

PR #1408 promotes latency-and-power overlays for these cohorts:

• B200 vLLM 0.22.0
• B200 TRT-LLM 1.3.0rc15
• H200 vLLM 0.22.0
• GB200 vLLM 0.22.0
• GB200 TRT-LLM 1.3.0rc17

PR #1408 also adds a conservative 90% latency-weighted energy-coverage gate. Latency, throughput, and memory estimates remain available below the threshold, while power_w is reported as unavailable.

A representative dense SILICON probe (Qwen/Qwen3-32B-FP8, TP=4, BS=8, ISL=2048, OSL=128) at PR #1408 head 0802f59f9c95c04155028decd734801daaf49ad0 measured:

Cohort	Power coverage	Result
B200 vLLM 0.22.0	95.21%	available
B200 TRT-LLM 1.3.0rc15	23.08%	unavailable
H200 vLLM 0.22.0	87.70%	unavailable
GB200 vLLM 0.22.0	97.46%	available
GB200 TRT-LLM 1.3.0rc17	74.22%	unavailable

Coverage is command-dependent; these values are representative probes, not cohort-wide guarantees.

The pre-#1358 MoE latency-inflation/recollection problem is separate and already tracked in #1414.

Work

• Inventory uncovered operations and shapes for representative dense and MoE configurations in every promoted cohort.
• Recollect or extend the missing power measurements.
• Publish fresh data through Collector V3 with full collection provenance and hashes.
• Verify representative dense and MoE commands meet the agreed coverage threshold, or explicitly document configurations that remain partial.
• Add data-driven end-to-end regression tests covering available and unavailable power outcomes for each promoted cohort.
• Validate calibration separately from completeness; passing the coverage gate must not be treated as an accuracy guarantee.

Acceptance criteria

• No command reports partial energy as a complete power estimate.
• Targeted supported dense and MoE configurations have at least 90% latency-weighted positive-energy coverage.
• Newly collected directories carry non-legacy Collector V3 provenance.
• Cohort-level regression coverage is committed and runs in CI.

Related: #1408, #1414