Integrate September Databricks results and consolidate full-path comparisons

[tom-clickhouse]

PR #42 added the September Databricks allocation exports. This change integrates those exports into the accepted real-time comparison, publishes the matched evidence and regenerated charts, and documents reproduction across all five systems.

Databricks preparation cost is $695.60475043, including Zerobus ingestion, MV refresh, and estimated Predictive Optimization. Adding $2.64835177 of normalized query cost gives $698.25310220 full-path cost and a 752.4914× cost-runtime product relative to the matched ClickHouse baseline. The accepted query selections remain 189 dashboard observations and 32 drill-down observations, or 820 query jobs per system.

Preparation costs use the supplied active-ingestion allocation window ending at producer completion. Query cost uses the accepted matched observations. Allocated usage, estimated maintenance, and normalized query cost retain their distinct provenance and pricing basis.

The Redshift update applies the documented, benchmark-owner-directed 32-active-RPU writer assumption. The original long-run measurements are retained; the model is explicitly distinguished from original-run billing evidence. SUPER and typed alternatives each reuse the same writer-plus-MSK path once.

Commits separate:

1. Redshift capacity assumption, cost summaries, and dependent charts.
2. Databricks allocation parsers, tests, matched observations, accepted summaries, and provider-pair charts.
3. Consolidated global charts, manifest integration, and current provenance.
4. Repository evidence map and portable reproduction instructions.

Validation:

• 109 Databricks offline tests passed.
• All 34 packaged Databricks artifacts verified against recorded size and SHA-256.
• Both cost command pipelines reproduced the accepted summaries without numerical changes.
• All 326 source-hash references in the affected derived summaries verified.
• 12 standard/wide CSV pairs match exactly.
• Six global latency/freshness charts regenerated with identical RGBA pixels and data; only their stale manifest hashes needed refreshing.
• Repository hygiene and source-text whitespace checks passed. Generated SVG/CSV serialization was excluded from the whitespace check.

No provider rerun is required. Historical stashes and the two unrelated June telemetry screenshots remain local and outside this PR.