A public write-up of 0.3: the benchmarks, the cost analysis, what nine crates cover

[sshaplygin]

Part of #63. Human-gated — nothing here happens without an explicit decision to post.

Nine crates, two transports and a benchmark against the C++ client, and no public account of any of it. The material exists; what is missing is a public write-up.

• the C++ comparison, with its own caveat kept in front of it: 1.66× less CPU on three columns and 1.26× on all 28, against column projection worth 5.9–7.8× — the honest headline is that what you read matters more than which client reads it
• what nine crates cover, and what they do not: four are pre-release in everything but their presence on the registry, and their gates are named in-tree
• where the RPC scope stops — transactions, lookup_rows, select_rows, modify_rows, and 154 other methods reachable only raw
• no claim the compatibility documents do not already support