[forhappy]

The Cloudflare benchmark returned second-scale reads and writes and roughly 26–29-second transaction medians at 60 clients. This investigation separates storage, metadata, coordinator activation and transaction protocol costs using the live three-node deployment and the unchanged executable.

• Add per-API findings, a reviewed aggregate CSV and an SVG showing medians, observed ranges and sample sizes.
• Direct fresh-object R2 writes take approximately one second through both S3 and the native binding. Native bucket-proof counters corroborate the foreground storage wait.
• Explain uncached metadata reads, read-tail correlation with background bucket proofs, durable transaction phase amplification and fresh coordinator activation costs. Separate confirmed code paths from remaining timing uncertainty.
• Retain diagnostic errors, small-sample limits and the restart-induced placement change (10/10/10 originally; 11/12/7 after diagnosis). No production code, dependency, authentication, durability or transaction semantics changed.

Validation: documentation/link/example/SVG checks; aggregate count/range checks; credential-value scan; visual inspection; diff whitespace check. Live checks confirmed all three signed endpoints, the requested platform allocations and all 30 uniquely owned resident data Cells. Strong readbacks verified 210/210 original values and 51/51 additional diagnostic write values. Temporary Worker probes and the R2 binding were removed; the original Worker and pinned image were restored. Failed/interrupted probes remain in private retained artifacts and the report; unattended recovery and production capacity remain unqualified.