eval: measure original SQL execution and deployment costs

[zzylol]

The original SQL workload needs execution and deployment measurements after automatic catalog construction, rather than counting successful bindings as acceleration.

This adds the original-corpus compiler, lossless capture importer, process replay drivers, raw-artifact hashes and a scoped report. Before this PR, the new typed Map path had process tests but no complete original-query matrix or fresh deployment comparison. After it, all 27 queries have explicit warm/fallback/failure outcomes on 12,236,002 real samples across 142 series; the three warm queries use actual Planner selections and installed SummaryStore state.

Three independent release trials per warm query contain 18,000 matched requests: all 9,000 backend requests are warm and decoded results equal native ClickHouse. Median speedups are 21.04–21.74× for q05, 8.37–8.97× for q06, and 13.84–14.70× for q23. Three separate q05 deployment pairs include startup, summary construction and 1,000 queries: backend plus retained ClickHouse uses 11.16–12.33 CPU seconds versus 96.67–98.86 seconds native, but combined RSS is higher. The report explicitly records that negative memory result.

Verification: importer tests (3), shell syntax checks, real automatic publication/backfill/query runs, and independent validation of all raw responses and 49 artifact hashes. Formal replay fails unless every requested repetition is warm and equal. Visual evidence is not applicable.

Limits: 3/27 warm, 18 correct whole-query fallbacks, and 6 native/fallback memory-limit failures; no broad SQL acceleration claim. Fixed evaluation windows, per-query publications, no numeric estimated-cost trace, and common initial ingestion/offline planning excluded. Exact source capture and raw traces remain outside Git with paths/hashes. Depends on #641.